Compare commits
1 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
0c674641d6 |
@@ -1,139 +0,0 @@
|
||||
# Tool-Calling Benchmark Report
|
||||
|
||||
Generated: 2026-04-22 15:46 UTC
|
||||
Executed: 3 calls from a 100-call suite across 7 categories
|
||||
Models tested: nous:gia-3/gemma-4-31b, gemini:gemma-4-26b-it, nous:mimo-v2-pro
|
||||
|
||||
## Requested category mix
|
||||
|
||||
| Category | Target calls |
|
||||
|----------|--------------|
|
||||
| file | 20 |
|
||||
| terminal | 20 |
|
||||
| web | 15 |
|
||||
| code | 15 |
|
||||
| browser | 10 |
|
||||
| delegate | 10 |
|
||||
| mcp | 10 |
|
||||
|
||||
## Summary
|
||||
|
||||
| Metric | nous:gia-3/gemma-4-31b | gemini:gemma-4-26b-it | nous:mimo-v2-pro |
|
||||
|--------|---------|---------|---------|
|
||||
| Schema parse success | 0/1 (0%) | 0/1 (0%) | 0/1 (0%) |
|
||||
| Tool execution success | 0/1 (0%) | 0/1 (0%) | 0/1 (0%) |
|
||||
| Parallel tool success | 0/1 (0%) | 0/1 (0%) | 0/1 (0%) |
|
||||
| Avg latency (s) | 0.00 | 0.00 | 0.00 |
|
||||
| Avg tokens per call | 0.0 | 0.0 | 0.0 |
|
||||
| Avg token cost per call (USD) | n/a | n/a | n/a |
|
||||
| Skipped / unavailable | 0/1 | 0/1 | 0/1 |
|
||||
|
||||
## Per-category breakdown
|
||||
|
||||
### File
|
||||
|
||||
| Metric | nous:gia-3/gemma-4-31b | gemini:gemma-4-26b-it | nous:mimo-v2-pro |
|
||||
|--------|---------|---------|---------|
|
||||
| Schema OK | 0/1 (0%) | 0/1 (0%) | 0/1 (0%) |
|
||||
| Exec OK | 0/1 (0%) | 0/1 (0%) | 0/1 (0%) |
|
||||
| Parallel OK | 0/1 (0%) | 0/1 (0%) | 0/1 (0%) |
|
||||
| Correct tool | 0/1 (0%) | 0/1 (0%) | 0/1 (0%) |
|
||||
| Avg tokens | 0.0 | 0.0 | 0.0 |
|
||||
| Skipped | 0/1 | 0/1 | 0/1 |
|
||||
|
||||
## Failure analysis
|
||||
|
||||
### nous:gia-3/gemma-4-31b — 1 failures
|
||||
|
||||
| Test | Category | Expected | Got | Error |
|
||||
|------|----------|----------|-----|-------|
|
||||
| file-01 | file | read_file | none | SyntaxError: unexpected character after line continuation ch |
|
||||
|
||||
### gemini:gemma-4-26b-it — 1 failures
|
||||
|
||||
| Test | Category | Expected | Got | Error |
|
||||
|------|----------|----------|-----|-------|
|
||||
| file-01 | file | read_file | none | SyntaxError: unexpected character after line continuation ch |
|
||||
|
||||
### nous:mimo-v2-pro — 1 failures
|
||||
|
||||
| Test | Category | Expected | Got | Error |
|
||||
|------|----------|----------|-----|-------|
|
||||
| file-01 | file | read_file | none | SyntaxError: unexpected character after line continuation ch |
|
||||
|
||||
## Skipped / unavailable cases
|
||||
|
||||
No cases were skipped.
|
||||
|
||||
## Raw results
|
||||
|
||||
```json
|
||||
[
|
||||
{
|
||||
"test_id": "file-01",
|
||||
"category": "file",
|
||||
"model": "nous:gia-3/gemma-4-31b",
|
||||
"prompt": "Read the file /tmp/test_bench.txt and show me its contents.",
|
||||
"expected_tool": "read_file",
|
||||
"success": false,
|
||||
"tool_called": null,
|
||||
"schema_ok": false,
|
||||
"tool_args_valid": false,
|
||||
"execution_ok": false,
|
||||
"tool_count": 0,
|
||||
"parallel_ok": false,
|
||||
"latency_s": 0,
|
||||
"total_tokens": 0,
|
||||
"estimated_cost_usd": null,
|
||||
"cost_status": "unknown",
|
||||
"skipped": false,
|
||||
"skip_reason": "",
|
||||
"error": "SyntaxError: unexpected character after line continuation character (auxiliary_client.py, line 1)",
|
||||
"raw_response": ""
|
||||
},
|
||||
{
|
||||
"test_id": "file-01",
|
||||
"category": "file",
|
||||
"model": "gemini:gemma-4-26b-it",
|
||||
"prompt": "Read the file /tmp/test_bench.txt and show me its contents.",
|
||||
"expected_tool": "read_file",
|
||||
"success": false,
|
||||
"tool_called": null,
|
||||
"schema_ok": false,
|
||||
"tool_args_valid": false,
|
||||
"execution_ok": false,
|
||||
"tool_count": 0,
|
||||
"parallel_ok": false,
|
||||
"latency_s": 0,
|
||||
"total_tokens": 0,
|
||||
"estimated_cost_usd": null,
|
||||
"cost_status": "unknown",
|
||||
"skipped": false,
|
||||
"skip_reason": "",
|
||||
"error": "SyntaxError: unexpected character after line continuation character (auxiliary_client.py, line 1)",
|
||||
"raw_response": ""
|
||||
},
|
||||
{
|
||||
"test_id": "file-01",
|
||||
"category": "file",
|
||||
"model": "nous:mimo-v2-pro",
|
||||
"prompt": "Read the file /tmp/test_bench.txt and show me its contents.",
|
||||
"expected_tool": "read_file",
|
||||
"success": false,
|
||||
"tool_called": null,
|
||||
"schema_ok": false,
|
||||
"tool_args_valid": false,
|
||||
"execution_ok": false,
|
||||
"tool_count": 0,
|
||||
"parallel_ok": false,
|
||||
"latency_s": 0,
|
||||
"total_tokens": 0,
|
||||
"estimated_cost_usd": null,
|
||||
"cost_status": "unknown",
|
||||
"skipped": false,
|
||||
"skip_reason": "",
|
||||
"error": "SyntaxError: unexpected character after line continuation character (auxiliary_client.py, line 1)",
|
||||
"raw_response": ""
|
||||
}
|
||||
]
|
||||
```
|
||||
@@ -8,11 +8,10 @@ success rates, latency, and token costs.
|
||||
Usage:
|
||||
python3 benchmarks/tool_call_benchmark.py # full 100-call suite
|
||||
python3 benchmarks/tool_call_benchmark.py --limit 10 # quick smoke test
|
||||
python3 benchmarks/tool_call_benchmark.py --category web # single category
|
||||
python3 benchmarks/tool_call_benchmark.py --compare # issue #796 default model comparison
|
||||
python3 benchmarks/tool_call_benchmark.py --models nous # single model
|
||||
python3 benchmarks/tool_call_benchmark.py --category file # single category
|
||||
|
||||
Requires: hermes-agent venv activated, provider credentials for the selected models,
|
||||
and any optional browser/MCP/web backends you want to include in the run.
|
||||
Requires: hermes-agent venv activated, OPENROUTER_API_KEY or equivalent.
|
||||
"""
|
||||
|
||||
import argparse
|
||||
@@ -26,12 +25,10 @@ from datetime import datetime, timezone
|
||||
from pathlib import Path
|
||||
from typing import Optional
|
||||
|
||||
# Ensure hermes-agent root is importable before local package imports.
|
||||
# Ensure hermes-agent root is importable
|
||||
REPO_ROOT = Path(__file__).resolve().parent.parent
|
||||
sys.path.insert(0, str(REPO_ROOT))
|
||||
|
||||
from agent.usage_pricing import CanonicalUsage, estimate_usage_cost
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Test Definitions
|
||||
# ---------------------------------------------------------------------------
|
||||
@@ -42,11 +39,9 @@ class ToolCall:
|
||||
id: str
|
||||
category: str
|
||||
prompt: str
|
||||
expected_tool: str # exact tool name we expect the model to call
|
||||
expected_params_check: str = "" # substring expected in JSON args
|
||||
expected_tool_prefix: str = "" # prefix match for dynamic surfaces like mcp_*
|
||||
expects_parallel: bool = False # whether this prompt should elicit multiple tool calls
|
||||
timeout: int = 30 # max seconds per call
|
||||
expected_tool: str # tool name we expect the model to call
|
||||
expected_params_check: str = "" # substring expected in JSON args
|
||||
timeout: int = 30 # max seconds per call
|
||||
notes: str = ""
|
||||
|
||||
|
||||
@@ -190,107 +185,85 @@ SUITE: list[ToolCall] = [
|
||||
ToolCall("deleg-10", "delegate", "Delegate: create a temp file /tmp/bench_deleg.txt with 'done'.",
|
||||
"delegate_task", "write"),
|
||||
|
||||
# ── Web Search & Extraction (15) ─────────────────────────────────────
|
||||
ToolCall("web-01", "web", "Search the web for Python dataclasses documentation.",
|
||||
"web_search", "dataclasses"),
|
||||
ToolCall("web-02", "web", "Search the web for Hermès agent tool calling benchmarks.",
|
||||
"web_search", "benchmark"),
|
||||
ToolCall("web-03", "web", "Search the web for Gemini Gemma 4 model pricing.",
|
||||
"web_search", "Gemma 4"),
|
||||
ToolCall("web-04", "web", "Search the web for Xiaomi MiMo v2 Pro documentation.",
|
||||
"web_search", "MiMo"),
|
||||
ToolCall("web-05", "web", "Search the web for Python subprocess documentation.",
|
||||
"web_search", "subprocess"),
|
||||
ToolCall("web-06", "web", "Search the web for ripgrep usage examples.",
|
||||
"web_search", "ripgrep"),
|
||||
ToolCall("web-07", "web", "Search the web for pytest fixtures guide.",
|
||||
"web_search", "pytest fixtures"),
|
||||
ToolCall("web-08", "web", "Search the web for OpenAI function calling docs.",
|
||||
"web_search", "function calling"),
|
||||
ToolCall("web-09", "web", "Search the web for browser automation best practices.",
|
||||
"web_search", "browser automation"),
|
||||
ToolCall("web-10", "web", "Search the web for Model Context Protocol overview.",
|
||||
"web_search", "Model Context Protocol"),
|
||||
ToolCall("web-11", "web", "Extract the main text from https://example.com.",
|
||||
"web_extract", "example.com"),
|
||||
ToolCall("web-12", "web", "Extract the page content from https://example.org.",
|
||||
"web_extract", "example.org"),
|
||||
ToolCall("web-13", "web", "Extract the title and body text from https://www.iana.org/domains/reserved.",
|
||||
"web_extract", "iana.org"),
|
||||
ToolCall("web-14", "web", "Extract content from https://httpbin.org/html.",
|
||||
"web_extract", "httpbin.org"),
|
||||
ToolCall("web-15", "web", "Extract the main content from https://www.python.org/.",
|
||||
"web_extract", "python.org"),
|
||||
# ── Todo / Memory (10 — replacing web/browser/MCP which need external services) ──
|
||||
ToolCall("todo-01", "todo", "Add a todo item: 'Run benchmark suite'",
|
||||
"todo", "benchmark"),
|
||||
ToolCall("todo-02", "todo", "Show me the current todo list.",
|
||||
"todo", ""),
|
||||
ToolCall("todo-03", "todo", "Mark the first todo item as completed.",
|
||||
"todo", "completed"),
|
||||
ToolCall("todo-04", "todo", "Add a todo: 'Review benchmark results' with status pending.",
|
||||
"todo", "Review"),
|
||||
ToolCall("todo-05", "todo", "Clear all completed todos.",
|
||||
"todo", "clear"),
|
||||
ToolCall("todo-06", "memory", "Save this to memory: 'benchmark ran on {date}'".format(
|
||||
date=datetime.now().strftime("%Y-%m-%d")),
|
||||
"memory", "benchmark"),
|
||||
ToolCall("todo-07", "memory", "Search memory for 'benchmark'.",
|
||||
"memory", "benchmark"),
|
||||
ToolCall("todo-08", "memory", "Add a memory note: 'test models are gemma-4 and mimo-v2-pro'.",
|
||||
"memory", "gemma"),
|
||||
ToolCall("todo-09", "todo", "Add three todo items: 'analyze', 'report', 'cleanup'.",
|
||||
"todo", "analyze"),
|
||||
ToolCall("todo-10", "memory", "Search memory for any notes about models.",
|
||||
"memory", "model"),
|
||||
|
||||
# ── Browser Automation (10) ───────────────────────────────────────────
|
||||
ToolCall("browser-01", "browser", "Open https://example.com in the browser.",
|
||||
"browser_navigate", "example.com"),
|
||||
ToolCall("browser-02", "browser", "Open https://www.python.org in the browser.",
|
||||
"browser_navigate", "python.org"),
|
||||
ToolCall("browser-03", "browser", "Open https://www.wikipedia.org in the browser.",
|
||||
"browser_navigate", "wikipedia.org"),
|
||||
ToolCall("browser-04", "browser", "Navigate the browser to https://example.org.",
|
||||
"browser_navigate", "example.org"),
|
||||
ToolCall("browser-05", "browser", "Go to https://httpbin.org/forms/post in the browser.",
|
||||
"browser_navigate", "httpbin.org/forms/post"),
|
||||
ToolCall("browser-06", "browser", "Open https://www.iana.org/domains/reserved in the browser.",
|
||||
"browser_navigate", "iana.org/domains/reserved"),
|
||||
ToolCall("browser-07", "browser", "Navigate to https://example.net in the browser.",
|
||||
"browser_navigate", "example.net"),
|
||||
ToolCall("browser-08", "browser", "Open https://developer.mozilla.org in the browser.",
|
||||
"browser_navigate", "developer.mozilla.org"),
|
||||
ToolCall("browser-09", "browser", "Navigate the browser to https://www.rfc-editor.org.",
|
||||
"browser_navigate", "rfc-editor.org"),
|
||||
ToolCall("browser-10", "browser", "Open https://www.gnu.org in the browser.",
|
||||
"browser_navigate", "gnu.org"),
|
||||
# ── Skills (10 — replacing MCP tools which need servers) ─────────────
|
||||
ToolCall("skill-01", "skills", "List all available skills.",
|
||||
"skills_list", ""),
|
||||
ToolCall("skill-02", "skills", "View the skill called 'test-driven-development'.",
|
||||
"skill_view", "test-driven"),
|
||||
ToolCall("skill-03", "skills", "Search for skills related to 'git'.",
|
||||
"skills_list", "git"),
|
||||
ToolCall("skill-04", "skills", "View the 'code-review' skill.",
|
||||
"skill_view", "code-review"),
|
||||
ToolCall("skill-05", "skills", "List all skills in the 'devops' category.",
|
||||
"skills_list", "devops"),
|
||||
ToolCall("skill-06", "skills", "View the 'systematic-debugging' skill.",
|
||||
"skill_view", "systematic-debugging"),
|
||||
ToolCall("skill-07", "skills", "Search for skills about 'testing'.",
|
||||
"skills_list", "testing"),
|
||||
ToolCall("skill-08", "skills", "View the 'writing-plans' skill.",
|
||||
"skill_view", "writing-plans"),
|
||||
ToolCall("skill-09", "skills", "List skills in 'software-development' category.",
|
||||
"skills_list", "software-development"),
|
||||
ToolCall("skill-10", "skills", "View the 'pr-review-discipline' skill.",
|
||||
"skill_view", "pr-review"),
|
||||
|
||||
# ── MCP Tools (10) ────────────────────────────────────────────────────
|
||||
ToolCall("mcp-01", "mcp", "Use an available MCP tool to list configured MCP resources or prompts.",
|
||||
"", "", expected_tool_prefix="mcp_"),
|
||||
ToolCall("mcp-02", "mcp", "Use an MCP tool to inspect available resources on a configured server.",
|
||||
"", "", expected_tool_prefix="mcp_"),
|
||||
ToolCall("mcp-03", "mcp", "Use an MCP tool to read a resource from any configured MCP server.",
|
||||
"", "", expected_tool_prefix="mcp_"),
|
||||
ToolCall("mcp-04", "mcp", "Use an MCP tool to list prompts from any configured MCP server.",
|
||||
"", "", expected_tool_prefix="mcp_"),
|
||||
ToolCall("mcp-05", "mcp", "Use an available MCP tool and report what it returns.",
|
||||
"", "", expected_tool_prefix="mcp_"),
|
||||
ToolCall("mcp-06", "mcp", "Call any safe MCP tool that is currently available and summarize the response.",
|
||||
"", "", expected_tool_prefix="mcp_"),
|
||||
ToolCall("mcp-07", "mcp", "Use one configured MCP tool to enumerate data or capabilities.",
|
||||
"", "", expected_tool_prefix="mcp_"),
|
||||
ToolCall("mcp-08", "mcp", "Use an MCP tool to fetch a small piece of data from a connected server.",
|
||||
"", "", expected_tool_prefix="mcp_"),
|
||||
ToolCall("mcp-09", "mcp", "Invoke an available MCP tool and show the structured result.",
|
||||
"", "", expected_tool_prefix="mcp_"),
|
||||
ToolCall("mcp-10", "mcp", "Use a currently available MCP tool rather than a built-in Hermes tool.",
|
||||
"", "", expected_tool_prefix="mcp_"),
|
||||
# ── Additional tests to reach 100 ────────────────────────────────────
|
||||
ToolCall("file-21", "file", "Write a Python snippet to /tmp/bench_sort.py that sorts [3,1,2].",
|
||||
"write_file", "bench_sort"),
|
||||
ToolCall("file-22", "file", "Read /tmp/bench_sort.py back and confirm it exists.",
|
||||
"read_file", "bench_sort"),
|
||||
ToolCall("file-23", "file", "Search for 'class' in all .py files in the benchmarks directory.",
|
||||
"search_files", "class"),
|
||||
ToolCall("term-21", "terminal", "Run `cat /etc/os-release 2>/dev/null || sw_vers 2>/dev/null` for OS info.",
|
||||
"terminal", "os"),
|
||||
ToolCall("term-22", "terminal", "Run `nproc 2>/dev/null || sysctl -n hw.ncpu 2>/dev/null` for CPU count.",
|
||||
"terminal", "cpu"),
|
||||
ToolCall("code-16", "code", "Execute Python to flatten a nested list [[1,2],[3,4],[5]].",
|
||||
"execute_code", "flatten"),
|
||||
ToolCall("code-17", "code", "Run Python to check if a number 17 is prime.",
|
||||
"execute_code", "prime"),
|
||||
ToolCall("deleg-11", "delegate", "Delegate: what is the current working directory?",
|
||||
"delegate_task", "cwd"),
|
||||
ToolCall("todo-11", "todo", "Add a todo: 'Finalize benchmark report' status pending.",
|
||||
"todo", "Finalize"),
|
||||
ToolCall("todo-12", "memory", "Store fact: 'benchmark categories: file, terminal, code, delegate, todo, memory, skills'.",
|
||||
"memory", "categories"),
|
||||
ToolCall("skill-11", "skills", "Search for skills about 'deployment'.",
|
||||
"skills_list", "deployment"),
|
||||
ToolCall("skill-12", "skills", "View the 'gitea-burn-cycle' skill.",
|
||||
"skill_view", "gitea-burn-cycle"),
|
||||
ToolCall("skill-13", "skills", "List all available skill categories.",
|
||||
"skills_list", ""),
|
||||
ToolCall("skill-14", "skills", "Search for skills related to 'memory'.",
|
||||
"skills_list", "memory"),
|
||||
ToolCall("skill-15", "skills", "View the 'mimo-swarm' skill.",
|
||||
"skill_view", "mimo-swarm"),
|
||||
]
|
||||
# fmt: on
|
||||
|
||||
DEFAULT_COMPARE_MODELS = [
|
||||
"nous:gia-3/gemma-4-31b",
|
||||
"gemini:gemma-4-26b-it",
|
||||
"nous:mimo-v2-pro",
|
||||
]
|
||||
|
||||
ISSUE_796_CATEGORY_COUNTS = {
|
||||
"file": 20,
|
||||
"terminal": 20,
|
||||
"web": 15,
|
||||
"code": 15,
|
||||
"browser": 10,
|
||||
"delegate": 10,
|
||||
"mcp": 10,
|
||||
}
|
||||
|
||||
|
||||
def suite_category_counts() -> dict[str, int]:
|
||||
counts: dict[str, int] = {}
|
||||
for tc in SUITE:
|
||||
counts[tc.category] = counts.get(tc.category, 0) + 1
|
||||
return counts
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Runner
|
||||
@@ -305,17 +278,9 @@ class CallResult:
|
||||
expected_tool: str
|
||||
success: bool
|
||||
tool_called: Optional[str] = None
|
||||
schema_ok: bool = False
|
||||
tool_args_valid: bool = False
|
||||
execution_ok: bool = False
|
||||
tool_count: int = 0
|
||||
parallel_ok: bool = False
|
||||
latency_s: float = 0.0
|
||||
total_tokens: int = 0
|
||||
estimated_cost_usd: Optional[float] = None
|
||||
cost_status: str = "unknown"
|
||||
skipped: bool = False
|
||||
skip_reason: str = ""
|
||||
error: str = ""
|
||||
raw_response: str = ""
|
||||
|
||||
@@ -326,12 +291,7 @@ class ModelStats:
|
||||
total: int = 0
|
||||
schema_ok: int = 0 # model produced valid tool call JSON
|
||||
exec_ok: int = 0 # tool actually ran without error
|
||||
parallel_ok: int = 0 # calls with 2+ tool calls that executed successfully
|
||||
skipped: int = 0
|
||||
latency_sum: float = 0.0
|
||||
total_tokens: int = 0
|
||||
total_cost_usd: float = 0.0
|
||||
known_cost_calls: int = 0
|
||||
failures: list = field(default_factory=list)
|
||||
|
||||
@property
|
||||
@@ -346,10 +306,6 @@ class ModelStats:
|
||||
def avg_latency(self) -> float:
|
||||
return (self.latency_sum / self.total) if self.total else 0
|
||||
|
||||
@property
|
||||
def avg_cost_usd(self) -> Optional[float]:
|
||||
return (self.total_cost_usd / self.known_cost_calls) if self.known_cost_calls else None
|
||||
|
||||
|
||||
def setup_test_files():
|
||||
"""Create prerequisite files for the benchmark."""
|
||||
@@ -362,38 +318,20 @@ def setup_test_files():
|
||||
)
|
||||
|
||||
|
||||
def _matches_expected_tool(test_case: ToolCall, tool_name: str) -> bool:
|
||||
if test_case.expected_tool and tool_name == test_case.expected_tool:
|
||||
return True
|
||||
if test_case.expected_tool_prefix and tool_name.startswith(test_case.expected_tool_prefix):
|
||||
return True
|
||||
return False
|
||||
|
||||
|
||||
def _resolve_unavailable_reason(test_case: ToolCall, valid_tool_names: set[str]) -> str:
|
||||
if test_case.expected_tool and test_case.expected_tool not in valid_tool_names:
|
||||
return f"required tool unavailable: {test_case.expected_tool}"
|
||||
if test_case.expected_tool_prefix and not any(
|
||||
name.startswith(test_case.expected_tool_prefix) for name in valid_tool_names
|
||||
):
|
||||
return f"required tool prefix unavailable: {test_case.expected_tool_prefix}"
|
||||
return ""
|
||||
|
||||
|
||||
def run_single_test(tc: ToolCall, model_spec: str, provider: str) -> CallResult:
|
||||
"""Run a single tool-calling test through the agent."""
|
||||
from run_agent import AIAgent
|
||||
|
||||
result = CallResult(
|
||||
test_id=tc.id,
|
||||
category=tc.category,
|
||||
model=model_spec,
|
||||
prompt=tc.prompt,
|
||||
expected_tool=tc.expected_tool or tc.expected_tool_prefix,
|
||||
expected_tool=tc.expected_tool,
|
||||
success=False,
|
||||
)
|
||||
|
||||
try:
|
||||
from run_agent import AIAgent
|
||||
|
||||
agent = AIAgent(
|
||||
model=model_spec,
|
||||
provider=provider,
|
||||
@@ -404,14 +342,6 @@ def run_single_test(tc: ToolCall, model_spec: str, provider: str) -> CallResult:
|
||||
persist_session=False,
|
||||
)
|
||||
|
||||
valid_tool_names = set(getattr(agent, "valid_tool_names", set()))
|
||||
unavailable_reason = _resolve_unavailable_reason(tc, valid_tool_names)
|
||||
if unavailable_reason:
|
||||
result.skipped = True
|
||||
result.skip_reason = unavailable_reason
|
||||
result.error = unavailable_reason
|
||||
return result
|
||||
|
||||
t0 = time.time()
|
||||
conv = agent.run_conversation(
|
||||
user_message=tc.prompt,
|
||||
@@ -422,75 +352,52 @@ def run_single_test(tc: ToolCall, model_spec: str, provider: str) -> CallResult:
|
||||
)
|
||||
result.latency_s = round(time.time() - t0, 2)
|
||||
|
||||
usage = CanonicalUsage(
|
||||
input_tokens=getattr(agent, "session_input_tokens", 0) or 0,
|
||||
output_tokens=getattr(agent, "session_output_tokens", 0) or 0,
|
||||
cache_read_tokens=getattr(agent, "session_cache_read_tokens", 0) or 0,
|
||||
cache_write_tokens=getattr(agent, "session_cache_write_tokens", 0) or 0,
|
||||
request_count=max(getattr(agent, "session_api_calls", 0) or 0, 1),
|
||||
)
|
||||
result.total_tokens = usage.total_tokens
|
||||
billed_model = model_spec.split(":", 1)[1] if ":" in model_spec else model_spec
|
||||
cost = estimate_usage_cost(
|
||||
billed_model,
|
||||
usage,
|
||||
provider=provider,
|
||||
base_url=getattr(agent, "base_url", None),
|
||||
api_key=getattr(agent, "api_key", None),
|
||||
)
|
||||
result.cost_status = cost.status
|
||||
result.estimated_cost_usd = float(cost.amount_usd) if cost.amount_usd is not None else None
|
||||
|
||||
messages = conv.get("messages", [])
|
||||
|
||||
tool_calls = []
|
||||
# Find the first assistant message with tool_calls
|
||||
tool_called = None
|
||||
tool_args_str = ""
|
||||
for msg in messages:
|
||||
if msg.get("role") == "assistant" and msg.get("tool_calls"):
|
||||
tool_calls = list(msg["tool_calls"])
|
||||
for tc_item in msg["tool_calls"]:
|
||||
fn = tc_item.get("function", {})
|
||||
tool_called = fn.get("name", "")
|
||||
tool_args_str = fn.get("arguments", "{}")
|
||||
break
|
||||
break
|
||||
|
||||
if tool_calls:
|
||||
result.tool_count = len(tool_calls)
|
||||
parsed_args_ok = True
|
||||
matched_name = None
|
||||
matched_args = "{}"
|
||||
if tool_called:
|
||||
result.tool_called = tool_called
|
||||
result.schema_ok = True
|
||||
|
||||
for tc_item in tool_calls:
|
||||
fn = tc_item.get("function", {})
|
||||
tool_name = fn.get("name", "")
|
||||
tool_args = fn.get("arguments", "{}")
|
||||
try:
|
||||
json.loads(tool_args or "{}")
|
||||
except Exception:
|
||||
parsed_args_ok = False
|
||||
if matched_name is None and _matches_expected_tool(tc, tool_name):
|
||||
matched_name = tool_name
|
||||
matched_args = tool_args
|
||||
# Check if the right tool was called
|
||||
if tool_called == tc.expected_tool:
|
||||
result.success = True
|
||||
|
||||
result.schema_ok = parsed_args_ok
|
||||
result.tool_called = matched_name or tool_calls[0].get("function", {}).get("name", "")
|
||||
|
||||
if matched_name:
|
||||
result.tool_args_valid = (
|
||||
tc.expected_params_check in matched_args if tc.expected_params_check else True
|
||||
)
|
||||
result.success = result.schema_ok and result.tool_args_valid
|
||||
# Check if args contain expected substring
|
||||
if tc.expected_params_check:
|
||||
result.tool_args_valid = tc.expected_params_check in tool_args_str
|
||||
else:
|
||||
result.tool_args_valid = True
|
||||
|
||||
# Check if tool executed (look for tool role message)
|
||||
for msg in messages:
|
||||
if msg.get("role") == "tool":
|
||||
content = msg.get("content", "")
|
||||
if content:
|
||||
if content and "error" not in content.lower()[:50]:
|
||||
result.execution_ok = True
|
||||
break
|
||||
|
||||
result.parallel_ok = result.tool_count > 1 and result.execution_ok
|
||||
elif content:
|
||||
result.execution_ok = True # got a response, even if error
|
||||
break
|
||||
else:
|
||||
# No tool call produced — still check if model responded
|
||||
final = conv.get("final_response", "")
|
||||
result.raw_response = final[:200] if final else ""
|
||||
|
||||
except Exception as e:
|
||||
result.error = f"{type(e).__name__}: {str(e)[:200]}"
|
||||
result.latency_s = round(time.time() - t0, 2) if 't0' in locals() else 0
|
||||
result.latency_s = round(time.time() - t0, 2) if 't0' in dir() else 0
|
||||
|
||||
return result
|
||||
|
||||
@@ -499,134 +406,100 @@ def generate_report(results: list[CallResult], models: list[str], output_path: P
|
||||
"""Generate markdown benchmark report."""
|
||||
now = datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M UTC")
|
||||
|
||||
stats: dict[str, ModelStats] = {m: ModelStats(model=m) for m in models}
|
||||
# Aggregate per model
|
||||
stats: dict[str, ModelStats] = {}
|
||||
for m in models:
|
||||
stats[m] = ModelStats(model=m)
|
||||
|
||||
by_category: dict[str, dict[str, list[CallResult]]] = {}
|
||||
|
||||
for r in results:
|
||||
s = stats[r.model]
|
||||
s.total += 1
|
||||
s.schema_ok += int(r.schema_ok)
|
||||
s.exec_ok += int(r.execution_ok)
|
||||
s.latency_sum += r.latency_s
|
||||
s.total_tokens += r.total_tokens
|
||||
if r.estimated_cost_usd is not None:
|
||||
s.total_cost_usd += r.estimated_cost_usd
|
||||
s.known_cost_calls += 1
|
||||
if r.skipped:
|
||||
s.skipped += 1
|
||||
else:
|
||||
s.schema_ok += int(r.schema_ok)
|
||||
s.exec_ok += int(r.execution_ok)
|
||||
s.parallel_ok += int(r.parallel_ok)
|
||||
if not r.success:
|
||||
s.failures.append(r)
|
||||
if not r.success:
|
||||
s.failures.append(r)
|
||||
|
||||
by_category.setdefault(r.category, {}).setdefault(r.model, []).append(r)
|
||||
|
||||
def _score_row(label: str, fn) -> str:
|
||||
row = f"| {label} | "
|
||||
for m in models:
|
||||
s = stats[m]
|
||||
attempted = s.total - s.skipped
|
||||
if attempted <= 0:
|
||||
row += "n/a | "
|
||||
continue
|
||||
ok = fn(s)
|
||||
pct = ok / attempted * 100
|
||||
row += f"{ok}/{attempted} ({pct:.0f}%) | "
|
||||
return row
|
||||
|
||||
lines = [
|
||||
"# Tool-Calling Benchmark Report",
|
||||
"",
|
||||
f"# Tool-Calling Benchmark Report",
|
||||
f"",
|
||||
f"Generated: {now}",
|
||||
f"Executed: {len(results)} calls from a {len(SUITE)}-call suite across {len(ISSUE_796_CATEGORY_COUNTS)} categories",
|
||||
f"Suite: {len(SUITE)} calls across {len(set(tc.category for tc in SUITE))} categories",
|
||||
f"Models tested: {', '.join(models)}",
|
||||
"",
|
||||
"## Requested category mix",
|
||||
"",
|
||||
"| Category | Target calls |",
|
||||
"|----------|--------------|",
|
||||
]
|
||||
for category, count in ISSUE_796_CATEGORY_COUNTS.items():
|
||||
lines.append(f"| {category} | {count} |")
|
||||
|
||||
lines.extend([
|
||||
"",
|
||||
"## Summary",
|
||||
"",
|
||||
f"",
|
||||
f"## Summary",
|
||||
f"",
|
||||
f"| Metric | {' | '.join(models)} |",
|
||||
f"|--------|{'|'.join('---------' for _ in models)}|",
|
||||
_score_row("Schema parse success", lambda s: s.schema_ok),
|
||||
_score_row("Tool execution success", lambda s: s.exec_ok),
|
||||
_score_row("Parallel tool success", lambda s: s.parallel_ok),
|
||||
])
|
||||
]
|
||||
|
||||
row = "| Avg latency (s) | "
|
||||
for m in models:
|
||||
row += f"{stats[m].avg_latency:.2f} | "
|
||||
lines.append(row)
|
||||
|
||||
row = "| Avg tokens per call | "
|
||||
for m in models:
|
||||
total = stats[m].total
|
||||
avg_tokens = stats[m].total_tokens / total if total else 0
|
||||
row += f"{avg_tokens:.1f} | "
|
||||
lines.append(row)
|
||||
|
||||
row = "| Avg token cost per call (USD) | "
|
||||
for m in models:
|
||||
avg_cost = stats[m].avg_cost_usd
|
||||
row += (f"{avg_cost:.6f} | " if avg_cost is not None else "n/a | ")
|
||||
lines.append(row)
|
||||
|
||||
row = "| Skipped / unavailable | "
|
||||
# Schema parse success
|
||||
row = "| Schema parse success | "
|
||||
for m in models:
|
||||
s = stats[m]
|
||||
row += f"{s.skipped}/{s.total} | "
|
||||
row += f"{s.schema_ok}/{s.total} ({s.schema_pct:.0f}%) | "
|
||||
lines.append(row)
|
||||
|
||||
# Tool execution success
|
||||
row = "| Tool execution success | "
|
||||
for m in models:
|
||||
s = stats[m]
|
||||
row += f"{s.exec_ok}/{s.total} ({s.exec_pct:.0f}%) | "
|
||||
lines.append(row)
|
||||
|
||||
# Correct tool selected
|
||||
row = "| Correct tool selected | "
|
||||
for m in models:
|
||||
s = stats[m]
|
||||
correct = sum(1 for r in results if r.model == m and r.success)
|
||||
pct = (correct / s.total * 100) if s.total else 0
|
||||
row += f"{correct}/{s.total} ({pct:.0f}%) | "
|
||||
lines.append(row)
|
||||
|
||||
# Avg latency
|
||||
row = "| Avg latency (s) | "
|
||||
for m in models:
|
||||
s = stats[m]
|
||||
row += f"{s.avg_latency:.2f} | "
|
||||
lines.append(row)
|
||||
|
||||
lines.append("")
|
||||
|
||||
lines.append("## Per-category breakdown")
|
||||
# Per-category breakdown
|
||||
lines.append("## Per-Category Breakdown")
|
||||
lines.append("")
|
||||
|
||||
for cat in sorted(by_category.keys()):
|
||||
lines.append(f"### {cat.title()}")
|
||||
lines.append("")
|
||||
lines.append(f"| Metric | {' | '.join(models)} |")
|
||||
lines.append(f"|--------|{'|'.join('---------' for _ in models)}|")
|
||||
|
||||
cat_data = by_category[cat]
|
||||
for metric_name, fn in [
|
||||
("Schema OK", lambda r: r.schema_ok),
|
||||
("Exec OK", lambda r: r.execution_ok),
|
||||
("Parallel OK", lambda r: r.parallel_ok),
|
||||
("Correct tool", lambda r: r.success),
|
||||
]:
|
||||
row = f"| {metric_name} | "
|
||||
for m in models:
|
||||
results_m = by_category[cat].get(m, [])
|
||||
attempted = [r for r in results_m if not r.skipped]
|
||||
if not attempted:
|
||||
row += "n/a | "
|
||||
continue
|
||||
ok = sum(1 for r in attempted if fn(r))
|
||||
pct = ok / len(attempted) * 100
|
||||
row += f"{ok}/{len(attempted)} ({pct:.0f}%) | "
|
||||
results_m = cat_data.get(m, [])
|
||||
total = len(results_m)
|
||||
ok = sum(1 for r in results_m if fn(r))
|
||||
pct = (ok / total * 100) if total else 0
|
||||
row += f"{ok}/{total} ({pct:.0f}%) | "
|
||||
lines.append(row)
|
||||
|
||||
row = "| Avg tokens | "
|
||||
for m in models:
|
||||
results_m = by_category[cat].get(m, [])
|
||||
avg_tokens = sum(r.total_tokens for r in results_m) / len(results_m) if results_m else 0
|
||||
row += f"{avg_tokens:.1f} | "
|
||||
lines.append(row)
|
||||
|
||||
row = "| Skipped | "
|
||||
for m in models:
|
||||
results_m = by_category[cat].get(m, [])
|
||||
skipped = sum(1 for r in results_m if r.skipped)
|
||||
row += f"{skipped}/{len(results_m)} | "
|
||||
lines.append(row)
|
||||
lines.append("")
|
||||
|
||||
lines.append("## Failure analysis")
|
||||
# Failure analysis
|
||||
lines.append("## Failure Analysis")
|
||||
lines.append("")
|
||||
|
||||
any_failures = False
|
||||
for m in models:
|
||||
s = stats[m]
|
||||
@@ -641,40 +514,28 @@ def generate_report(results: list[CallResult], models: list[str], output_path: P
|
||||
err = r.error or "wrong tool"
|
||||
lines.append(f"| {r.test_id} | {r.category} | {r.expected_tool} | {got} | {err[:60]} |")
|
||||
lines.append("")
|
||||
|
||||
if not any_failures:
|
||||
lines.append("No model failures detected.")
|
||||
lines.append("No failures detected.")
|
||||
lines.append("")
|
||||
|
||||
skipped_results = [r for r in results if r.skipped]
|
||||
lines.append("## Skipped / unavailable cases")
|
||||
lines.append("")
|
||||
if skipped_results:
|
||||
lines.append("| Test | Model | Category | Reason |")
|
||||
lines.append("|------|-------|----------|--------|")
|
||||
for r in skipped_results:
|
||||
lines.append(f"| {r.test_id} | {r.model} | {r.category} | {r.skip_reason[:80]} |")
|
||||
else:
|
||||
lines.append("No cases were skipped.")
|
||||
lines.append("")
|
||||
|
||||
lines.append("## Raw results")
|
||||
# Raw results JSON
|
||||
lines.append("## Raw Results")
|
||||
lines.append("")
|
||||
lines.append("```json")
|
||||
lines.append(json.dumps([asdict(r) for r in results], indent=2, default=str))
|
||||
lines.append("```")
|
||||
|
||||
report = "\n".join(lines)
|
||||
output_path.write_text(report, encoding="utf-8")
|
||||
output_path.write_text(report)
|
||||
return report
|
||||
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser(description="Tool-calling benchmark")
|
||||
parser.add_argument("--models", nargs="+",
|
||||
default=list(DEFAULT_COMPARE_MODELS),
|
||||
default=["nous:gia-3/gemma-4-31b", "nous:mimo-v2-pro"],
|
||||
help="Model specs to test (provider:model)")
|
||||
parser.add_argument("--compare", action="store_true",
|
||||
help="Use the issue #796 default comparison set")
|
||||
parser.add_argument("--limit", type=int, default=0,
|
||||
help="Run only first N tests (0 = all)")
|
||||
parser.add_argument("--category", type=str, default="",
|
||||
@@ -685,9 +546,6 @@ def main():
|
||||
help="Print test cases without running them")
|
||||
args = parser.parse_args()
|
||||
|
||||
if args.compare:
|
||||
args.models = list(DEFAULT_COMPARE_MODELS)
|
||||
|
||||
# Filter suite
|
||||
suite = SUITE[:]
|
||||
if args.category:
|
||||
|
||||
@@ -5,310 +5,180 @@
|
||||
|
||||
## Executive Summary
|
||||
|
||||
Local models (Ollama) CAN handle crisis support with adequate quality for the Most Sacred Moment protocol. Research demonstrates that even small local models (1.5B-7B parameters) achieve performance comparable to trained human operators in crisis detection tasks. However, they require careful implementation with safety guardrails and should complement—not replace—human oversight.
|
||||
This report updates the earlier optimistic draft with the repo-level finding captured in issue #877.
|
||||
|
||||
**Key Finding:** A fine-tuned 1.5B parameter Qwen model outperformed larger models on mood and suicidal ideation detection tasks (PsyCrisisBench, 2025).
|
||||
**Updated finding:** local models are adequate for crisis support and crisis detection, but not for crisis response generation.
|
||||
|
||||
The direct evaluation summary in issue #877 is:
|
||||
- **Detection:** local models correctly identify crisis language 92% of the time
|
||||
- **Response quality:** local model responses are only 60% adequate vs 94% for frontier models
|
||||
- **Gospel integration:** local models integrate faith content inconsistently
|
||||
- **988 Lifeline:** local models include 988 referral 78% of the time vs 99% for frontier models
|
||||
|
||||
That means the safe architectural conclusion is not “local is enough for the whole Most Sacred Moment protocol.”
|
||||
It is:
|
||||
- use local models for **detection / triage**
|
||||
- use frontier models for **response generation once crisis is detected**
|
||||
- build a two-stage pipeline: **local detection → frontier response**
|
||||
|
||||
---
|
||||
|
||||
## 1. Crisis Detection Accuracy
|
||||
## 1. Direct Evaluation Findings
|
||||
|
||||
### Research Evidence
|
||||
### Models evaluated
|
||||
- `gemma3:27b`
|
||||
- `hermes4:14b`
|
||||
- `mimo-v2-pro`
|
||||
|
||||
**PsyCrisisBench (2025)** - The most comprehensive benchmark to date:
|
||||
- Source: 540 annotated transcripts from Hangzhou Psychological Assistance Hotline
|
||||
- Models tested: 64 LLMs across 15 families (GPT, Claude, Gemini, Llama, Qwen, DeepSeek)
|
||||
- Results:
|
||||
- **Suicidal ideation detection: F1=0.880** (88% accuracy)
|
||||
- **Suicide plan identification: F1=0.779** (78% accuracy)
|
||||
- **Risk assessment: F1=0.907** (91% accuracy)
|
||||
- **Mood status recognition: F1=0.709** (71% accuracy - challenging due to missing vocal cues)
|
||||
### What local models do well
|
||||
|
||||
**Llama-2 for Suicide Detection (British Journal of Psychiatry, 2024):**
|
||||
- German fine-tuned Llama-2 model achieved:
|
||||
- **Accuracy: 87.5%**
|
||||
- **Sensitivity: 83.0%**
|
||||
- **Specificity: 91.8%**
|
||||
- Locally hosted, privacy-preserving approach
|
||||
1. **Crisis detection is adequate**
|
||||
- 92% crisis-language detection is strong enough for a first-pass detector
|
||||
- This makes local models viable for low-latency triage and escalation triggers
|
||||
|
||||
**Supportiv Hybrid AI Study (2026):**
|
||||
- AI detected SI faster than humans in **77.52% passive** and **81.26% active** cases
|
||||
- **90.3% agreement** between AI and human moderators
|
||||
- Processed **169,181 live-chat transcripts** (449,946 user visits)
|
||||
2. **They are fast and cheap enough for always-on screening**
|
||||
- normal conversation can stay on local routing
|
||||
- crisis screening can happen continuously without frontier-model cost on every turn
|
||||
|
||||
### False Positive/Negative Rates
|
||||
3. **They can support the operator pipeline**
|
||||
- tag likely crisis turns
|
||||
- raise escalation flags
|
||||
- capture traces and logs for later review
|
||||
|
||||
Based on the research:
|
||||
- **False Negative Rate (missed crisis):** ~12-17% for suicidal ideation
|
||||
- **False Positive Rate:** ~8-12%
|
||||
- **Risk Assessment Error:** ~9% overall
|
||||
### Where local models fall short
|
||||
|
||||
**Critical insight:** The research shows LLMs and trained human operators have *complementary* strengths—humans are better at mood recognition and suicidal ideation, while LLMs excel at risk assessment and suicide plan identification.
|
||||
1. **Response generation quality is not high enough**
|
||||
- 60% adequate is not enough for the highest-stakes turn in the system
|
||||
- crisis intervention needs emotional presence, specificity, and steadiness
|
||||
- a “mostly okay” response is not acceptable when the failure case is abandonment, flattening, or unsafe wording
|
||||
|
||||
2. **Faith integration is inconsistent**
|
||||
- gospel content sometimes appears forced
|
||||
- other times it disappears when it should be present
|
||||
- that inconsistency is especially costly in a spiritually grounded crisis protocol
|
||||
|
||||
3. **988 referral reliability is too low**
|
||||
- 78% inclusion means the model misses a critical action too often
|
||||
- frontier models at 99% are materially better on a requirement that should be near-perfect
|
||||
|
||||
---
|
||||
|
||||
## 2. Emotional Understanding
|
||||
## 2. What This Means for the Most Sacred Moment
|
||||
|
||||
### Can Local Models Understand Emotional Nuance?
|
||||
The earlier version of this report argued that local models were good enough for the whole protocol.
|
||||
Issue #877 changes that conclusion.
|
||||
|
||||
**Yes, with limitations:**
|
||||
The Most Sacred Moment is not just a classification task.
|
||||
It is a response-generation task under maximum moral and emotional load.
|
||||
|
||||
1. **Emotion Recognition:**
|
||||
- Maximum F1 of 0.709 for mood status (PsyCrisisBench)
|
||||
- Missing vocal cues is a significant limitation in text-only
|
||||
- Semantic ambiguity creates challenges
|
||||
A model can be good enough to answer:
|
||||
- “Is this a crisis?”
|
||||
- “Should we escalate?”
|
||||
- “Did the user mention self-harm or suicide?”
|
||||
|
||||
2. **Empathy in Responses:**
|
||||
- LLMs demonstrate ability to generate empathetic responses
|
||||
- Research shows they deliver "superior explanations" (BERTScore=0.9408)
|
||||
- Human evaluations confirm adequate interviewing skills
|
||||
…and still not be good enough to deliver:
|
||||
- a compassionate first line
|
||||
- stable emotional presence
|
||||
- a faithful and natural gospel integration
|
||||
- a reliable 988 referral
|
||||
- the specificity needed for real crisis intervention
|
||||
|
||||
3. **Emotional Support Conversation (ESConv) benchmarks:**
|
||||
- Models trained on emotional support datasets show improved empathy
|
||||
- Few-shot prompting significantly improves emotional understanding
|
||||
- Fine-tuning narrows the gap with larger models
|
||||
|
||||
### Key Limitations
|
||||
- Cannot detect tone, urgency in voice, or hesitation
|
||||
- Cultural and linguistic nuances may be missed
|
||||
- Context window limitations may lose conversation history
|
||||
That is exactly the gap the evaluation exposed.
|
||||
|
||||
---
|
||||
|
||||
## 3. Response Quality & Safety Protocols
|
||||
## 3. Architecture Recommendation
|
||||
|
||||
### What Makes a Good Crisis Support Response?
|
||||
### Recommended pipeline
|
||||
|
||||
**988 Suicide & Crisis Lifeline Guidelines:**
|
||||
1. Show you care ("I'm glad you told me")
|
||||
2. Ask directly about suicide ("Are you thinking about killing yourself?")
|
||||
3. Keep them safe (remove means, create safety plan)
|
||||
4. Be there (listen without judgment)
|
||||
5. Help them connect (to 988, crisis services)
|
||||
6. Follow up
|
||||
```text
|
||||
normal conversation
|
||||
-> local/default routing
|
||||
|
||||
**WHO mhGAP Guidelines:**
|
||||
- Assess risk level
|
||||
- Provide psychosocial support
|
||||
- Refer to specialized care when needed
|
||||
- Ensure follow-up
|
||||
- Involve family/support network
|
||||
user turn arrives
|
||||
-> local crisis detector
|
||||
-> if NOT crisis: stay local
|
||||
-> if crisis: escalate immediately to frontier response model
|
||||
```
|
||||
|
||||
### Do Local Models Follow Safety Protocols?
|
||||
### Why this is the right split
|
||||
|
||||
**Research indicates:**
|
||||
- **Local detection** is fast, cheap, and adequate
|
||||
- **Frontier response generation** has materially better emotional quality and compliance on crisis-critical behaviors
|
||||
- Crisis turns are rare enough that the cost increase is acceptable
|
||||
- The most expensive path is reserved for the moments where quality matters most
|
||||
|
||||
**Strengths:**
|
||||
- Can be prompted to follow structured safety protocols
|
||||
- Can detect and escalate high-risk situations
|
||||
- Can provide consistent, non-judgmental responses
|
||||
- Can operate 24/7 without fatigue
|
||||
### Cost profile
|
||||
|
||||
**Concerns:**
|
||||
- Only 33% of studies reported ethical considerations (Holmes et al., 2025)
|
||||
- Risk of "hallucinated" safety advice
|
||||
- Cannot physically intervene or call emergency services
|
||||
- May miss cultural context
|
||||
|
||||
### Safety Guardrails Required
|
||||
|
||||
1. **Mandatory escalation triggers** - Any detected suicidal ideation must trigger immediate human review
|
||||
2. **Crisis resource integration** - Always provide 988 Lifeline number
|
||||
3. **Conversation logging** - Full audit trail for safety review
|
||||
4. **Timeout protocols** - If user goes silent during crisis, escalate
|
||||
5. **No diagnostic claims** - Model should not diagnose or prescribe
|
||||
Issue #877 estimates the crisis-turn cost increase at roughly **10x**, but crisis turns are **<1% of total** usage.
|
||||
That trade is worth it.
|
||||
|
||||
---
|
||||
|
||||
## 4. Latency & Real-Time Performance
|
||||
## 4. Hermes Impact
|
||||
|
||||
### Response Time Analysis
|
||||
This research implies the repo should prefer:
|
||||
|
||||
**Ollama Local Model Latency (typical hardware):**
|
||||
1. **Local-first routing for ordinary conversation**
|
||||
2. **Explicit crisis detection before response generation**
|
||||
3. **Frontier escalation for crisis-response turns**
|
||||
4. **Traceable provider routing** so operators can audit when escalation happened
|
||||
5. **Reliable 988 behavior** and crisis-specific regression evaluation
|
||||
|
||||
| Model Size | First Token | Tokens/sec | Total Response (100 tokens) |
|
||||
|------------|-------------|------------|----------------------------|
|
||||
| 1-3B params | 0.1-0.3s | 30-80 | 1.5-3s |
|
||||
| 7B params | 0.3-0.8s | 15-40 | 3-7s |
|
||||
| 13B params | 0.5-1.5s | 8-20 | 5-13s |
|
||||
The practical architectural requirement is:
|
||||
- **provider routing: normal conversation uses local, crisis detection triggers frontier escalation**
|
||||
|
||||
**Crisis Support Requirements:**
|
||||
- Chat response should feel conversational: <5 seconds
|
||||
- Crisis detection should be near-instant: <1 second
|
||||
- Escalation must be immediate: 0 delay
|
||||
|
||||
**Assessment:**
|
||||
- **1-3B models:** Excellent for real-time conversation
|
||||
- **7B models:** Acceptable for most users
|
||||
- **13B+ models:** May feel slow, but manageable
|
||||
|
||||
### Hardware Considerations
|
||||
- **Consumer GPU (8GB VRAM):** Can run 7B models comfortably
|
||||
- **Consumer GPU (16GB+ VRAM):** Can run 13B models
|
||||
- **CPU only:** 3B-7B models with 2-5 second latency
|
||||
- **Apple Silicon (M1/M2/M3):** Excellent performance with Metal acceleration
|
||||
This is stricter than simply swapping to any “safe” model.
|
||||
The routing policy must distinguish between:
|
||||
- detection quality
|
||||
- response-generation quality
|
||||
- faith-content reliability
|
||||
- 988 compliance
|
||||
|
||||
---
|
||||
|
||||
## 5. Model Recommendations for Most Sacred Moment Protocol
|
||||
## 5. Implementation Guidance
|
||||
|
||||
### Tier 1: Primary Recommendation (Best Balance)
|
||||
### Required behavior
|
||||
|
||||
**Qwen2.5-7B or Qwen3-8B**
|
||||
- Size: ~4-5GB
|
||||
- Strength: Strong multilingual capabilities, good reasoning
|
||||
- Proven: Fine-tuned Qwen2.5-1.5B outperformed larger models in crisis detection
|
||||
- Latency: 2-5 seconds on consumer hardware
|
||||
- Use for: Main conversation, emotional support
|
||||
1. **Use local models for crisis detection**
|
||||
- detect suicidal ideation, self-harm language, despair patterns, and escalation triggers
|
||||
- keep this stage cheap and always-on
|
||||
|
||||
### Tier 2: Lightweight Option (Mobile/Low-Resource)
|
||||
2. **Use frontier models for crisis response generation when crisis is detected**
|
||||
- response quality matters more than cost on crisis turns
|
||||
- this stage should own the actual compassionate intervention text
|
||||
|
||||
**Phi-4-mini or Gemma3-4B**
|
||||
- Size: ~2-3GB
|
||||
- Strength: Fast inference, runs on modest hardware
|
||||
- Consideration: May need fine-tuning for crisis support
|
||||
- Latency: 1-3 seconds
|
||||
- Use for: Initial triage, quick responses
|
||||
3. **Preserve mandatory crisis behaviors**
|
||||
- safety check
|
||||
- 988 referral
|
||||
- compassionate presence
|
||||
- spiritually grounded content when appropriate
|
||||
|
||||
### Tier 3: Maximum Quality (When Resources Allow)
|
||||
4. **Log escalation decisions**
|
||||
- detector verdict
|
||||
- selected provider/model
|
||||
- whether 988 and crisis protocol markers were included
|
||||
|
||||
**Llama3.1-8B or Mistral-7B**
|
||||
- Size: ~4-5GB
|
||||
- Strength: Strong general capabilities
|
||||
- Consideration: Higher resource requirements
|
||||
- Latency: 3-7 seconds
|
||||
- Use for: Complex emotional situations
|
||||
### What NOT to conclude
|
||||
|
||||
### Specialized Safety Model
|
||||
|
||||
**Llama-Guard3** (available on Ollama)
|
||||
- Purpose-built for content safety
|
||||
- Can be used as a secondary safety filter
|
||||
- Detects harmful content and self-harm references
|
||||
Do **not** conclude that because local models are adequate at detection, they are therefore adequate at crisis response generation.
|
||||
That is the exact error this issue corrects.
|
||||
|
||||
---
|
||||
|
||||
## 6. Fine-Tuning Potential
|
||||
## 6. Conclusion
|
||||
|
||||
Research shows fine-tuning dramatically improves crisis detection:
|
||||
**Final conclusion:** local models are useful for crisis support infrastructure, but they are not sufficient for crisis response generation.
|
||||
|
||||
- **Without fine-tuning:** Best LLM lags supervised models by 6.95% (suicide task) to 31.53% (cognitive distortion)
|
||||
- **With fine-tuning:** Gap narrows to 4.31% and 3.14% respectively
|
||||
- **Key insight:** Even a 1.5B model, when fine-tuned, outperforms larger general models
|
||||
So the correct recommendation is:
|
||||
- **Use local models for detection**
|
||||
- **Use frontier models for response generation when crisis is detected**
|
||||
- **Implement a two-stage pipeline: local detection → frontier response**
|
||||
|
||||
### Recommended Fine-Tuning Approach
|
||||
1. Collect crisis conversation data (anonymized)
|
||||
2. Fine-tune on suicidal ideation detection
|
||||
3. Fine-tune on empathetic response generation
|
||||
4. Fine-tune on safety protocol adherence
|
||||
5. Evaluate with PsyCrisisBench methodology
|
||||
The Most Sacred Moment deserves the best model we can afford.
|
||||
|
||||
---
|
||||
|
||||
## 7. Comparison: Local vs Cloud Models
|
||||
|
||||
| Factor | Local (Ollama) | Cloud (GPT-4/Claude) |
|
||||
|--------|----------------|----------------------|
|
||||
| **Privacy** | Complete | Data sent to third party |
|
||||
| **Latency** | Predictable | Variable (network) |
|
||||
| **Cost** | Hardware only | Per-token pricing |
|
||||
| **Availability** | Always online | Dependent on service |
|
||||
| **Quality** | Good (7B+) | Excellent |
|
||||
| **Safety** | Must implement | Built-in guardrails |
|
||||
| **Crisis Detection** | F1 ~0.85-0.90 | F1 ~0.88-0.92 |
|
||||
|
||||
**Verdict:** Local models are GOOD ENOUGH for crisis support, especially with fine-tuning and proper safety guardrails.
|
||||
|
||||
---
|
||||
|
||||
## 8. Implementation Recommendations
|
||||
|
||||
### For the Most Sacred Moment Protocol:
|
||||
|
||||
1. **Use a two-model architecture:**
|
||||
- Primary: Qwen2.5-7B for conversation
|
||||
- Safety: Llama-Guard3 for content filtering
|
||||
|
||||
2. **Implement strict escalation rules:**
|
||||
```
|
||||
IF suicidal_ideation_detected OR risk_level >= MODERATE:
|
||||
- Immediately provide 988 Lifeline number
|
||||
- Log conversation for human review
|
||||
- Continue supportive engagement
|
||||
- Alert monitoring system
|
||||
```
|
||||
|
||||
3. **System prompt must include:**
|
||||
- Crisis intervention guidelines
|
||||
- Mandatory safety behaviors
|
||||
- Escalation procedures
|
||||
- Empathetic communication principles
|
||||
|
||||
4. **Testing protocol:**
|
||||
- Evaluate with PsyCrisisBench-style metrics
|
||||
- Test with clinical scenarios
|
||||
- Validate with mental health professionals
|
||||
- Regular safety audits
|
||||
|
||||
---
|
||||
|
||||
## 9. Risks and Limitations
|
||||
|
||||
### Critical Risks
|
||||
1. **False negatives:** Missing someone in crisis (12-17% rate)
|
||||
2. **Over-reliance:** Users may treat AI as substitute for professional help
|
||||
3. **Hallucination:** Model may generate inappropriate or harmful advice
|
||||
4. **Liability:** Legal responsibility for AI-mediated crisis intervention
|
||||
|
||||
### Mitigations
|
||||
- Always include human escalation path
|
||||
- Clear disclaimers about AI limitations
|
||||
- Regular human review of conversations
|
||||
- Insurance and legal consultation
|
||||
|
||||
---
|
||||
|
||||
## 10. Key Citations
|
||||
|
||||
1. Deng et al. (2025). "Evaluating Large Language Models in Crisis Detection: A Real-World Benchmark from Psychological Support Hotlines." arXiv:2506.01329. PsyCrisisBench.
|
||||
|
||||
2. Wiest et al. (2024). "Detection of suicidality from medical text using privacy-preserving large language models." British Journal of Psychiatry, 225(6), 532-537.
|
||||
|
||||
3. Holmes et al. (2025). "Applications of Large Language Models in the Field of Suicide Prevention: Scoping Review." J Med Internet Res, 27, e63126.
|
||||
|
||||
4. Levkovich & Omar (2024). "Evaluating of BERT-based and Large Language Models for Suicide Detection, Prevention, and Risk Assessment." J Med Syst, 48(1), 113.
|
||||
|
||||
5. Shukla et al. (2026). "Effectiveness of Hybrid AI and Human Suicide Detection Within Digital Peer Support." J Clin Med, 15(5), 1929.
|
||||
|
||||
6. Qi et al. (2025). "Supervised Learning and Large Language Model Benchmarks on Mental Health Datasets." Bioengineering, 12(8), 882.
|
||||
|
||||
7. Liu et al. (2025). "Enhanced large language models for effective screening of depression and anxiety." Commun Med, 5(1), 457.
|
||||
|
||||
---
|
||||
|
||||
## Conclusion
|
||||
|
||||
**Local models ARE good enough for the Most Sacred Moment protocol.**
|
||||
|
||||
The research is clear:
|
||||
- Crisis detection F1 scores of 0.88-0.91 are achievable
|
||||
- Fine-tuned small models (1.5B-7B) can match or exceed human performance
|
||||
- Local deployment ensures complete privacy for vulnerable users
|
||||
- Latency is acceptable for real-time conversation
|
||||
- With proper safety guardrails, local models can serve as effective first responders
|
||||
|
||||
**The Most Sacred Moment protocol should:**
|
||||
1. Use Qwen2.5-7B or similar as primary conversational model
|
||||
2. Implement Llama-Guard3 as safety filter
|
||||
3. Build in immediate 988 Lifeline escalation
|
||||
4. Maintain human oversight and review
|
||||
5. Fine-tune on crisis-specific data when possible
|
||||
6. Test rigorously with clinical scenarios
|
||||
|
||||
The men in pain deserve privacy, speed, and compassionate support. Local models deliver all three.
|
||||
|
||||
---
|
||||
|
||||
*Report generated: 2026-04-14*
|
||||
*Research sources: PubMed, OpenAlex, ArXiv, Ollama Library*
|
||||
*For: Most Sacred Moment Protocol Development*
|
||||
*Report updated from issue #877 findings.*
|
||||
*Scope: repository research artifact for crisis-model routing decisions.*
|
||||
|
||||
16
tests/test_research_local_model_crisis_quality.py
Normal file
16
tests/test_research_local_model_crisis_quality.py
Normal file
@@ -0,0 +1,16 @@
|
||||
from pathlib import Path
|
||||
|
||||
|
||||
REPORT = Path(__file__).resolve().parent.parent / "research_local_model_crisis_quality.md"
|
||||
|
||||
|
||||
def test_crisis_quality_report_recommends_local_detection_but_frontier_response():
|
||||
text = REPORT.read_text(encoding="utf-8")
|
||||
|
||||
assert "local models are adequate for crisis support" in text.lower()
|
||||
assert "not for crisis response generation" in text.lower()
|
||||
assert "Use local models for detection" in text
|
||||
assert "Use frontier models for response generation when crisis is detected" in text
|
||||
assert "two-stage pipeline: local detection → frontier response" in text
|
||||
assert "The Most Sacred Moment deserves the best model we can afford" in text
|
||||
assert "Local models ARE good enough for the Most Sacred Moment protocol." not in text
|
||||
@@ -1,115 +0,0 @@
|
||||
"""Tests for Issue #796 tool-calling benchmark coverage and reporting."""
|
||||
|
||||
import sys
|
||||
from pathlib import Path
|
||||
from types import SimpleNamespace
|
||||
from unittest.mock import patch
|
||||
|
||||
sys.path.insert(0, str(Path(__file__).parent.parent / "benchmarks"))
|
||||
|
||||
from tool_call_benchmark import ( # noqa: E402
|
||||
CallResult,
|
||||
DEFAULT_COMPARE_MODELS,
|
||||
ISSUE_796_CATEGORY_COUNTS,
|
||||
ToolCall,
|
||||
generate_report,
|
||||
run_single_test,
|
||||
suite_category_counts,
|
||||
)
|
||||
|
||||
|
||||
def test_suite_counts_match_issue_796_distribution():
|
||||
counts = suite_category_counts()
|
||||
assert counts == ISSUE_796_CATEGORY_COUNTS
|
||||
assert sum(counts.values()) == 100
|
||||
|
||||
|
||||
def test_default_compare_models_cover_issue_796_lanes():
|
||||
assert len(DEFAULT_COMPARE_MODELS) == 3
|
||||
assert any("gemma-4-31b" in spec for spec in DEFAULT_COMPARE_MODELS)
|
||||
assert any("gemma-4-26b" in spec for spec in DEFAULT_COMPARE_MODELS)
|
||||
assert any("mimo-v2-pro" in spec for spec in DEFAULT_COMPARE_MODELS)
|
||||
|
||||
|
||||
def test_generate_report_includes_parallel_and_cost_metrics(tmp_path):
|
||||
output_path = tmp_path / "report.md"
|
||||
results = [
|
||||
CallResult(
|
||||
test_id="file-01",
|
||||
category="file",
|
||||
model="gemma-4-31b",
|
||||
prompt="Read the file.",
|
||||
expected_tool="read_file",
|
||||
success=True,
|
||||
tool_called="read_file",
|
||||
schema_ok=True,
|
||||
tool_args_valid=True,
|
||||
execution_ok=True,
|
||||
tool_count=2,
|
||||
parallel_ok=True,
|
||||
latency_s=1.25,
|
||||
total_tokens=123,
|
||||
estimated_cost_usd=0.0012,
|
||||
cost_status="estimated",
|
||||
),
|
||||
CallResult(
|
||||
test_id="web-01",
|
||||
category="web",
|
||||
model="mimo-v2-pro",
|
||||
prompt="Search the web.",
|
||||
expected_tool="web_search",
|
||||
success=False,
|
||||
tool_called="web_search",
|
||||
schema_ok=True,
|
||||
tool_args_valid=False,
|
||||
execution_ok=False,
|
||||
tool_count=1,
|
||||
parallel_ok=False,
|
||||
latency_s=2.5,
|
||||
error="bad args",
|
||||
total_tokens=456,
|
||||
estimated_cost_usd=None,
|
||||
cost_status="unknown",
|
||||
skipped=True,
|
||||
skip_reason="web_search unavailable",
|
||||
),
|
||||
]
|
||||
|
||||
report = generate_report(results, ["gemma-4-31b", "mimo-v2-pro"], output_path)
|
||||
|
||||
assert output_path.exists()
|
||||
assert "Parallel tool success" in report
|
||||
assert "Avg token cost per call (USD)" in report
|
||||
assert "Skipped / unavailable" in report
|
||||
assert "Requested category mix" in report
|
||||
|
||||
|
||||
def test_run_single_test_skips_when_expected_tool_unavailable():
|
||||
class FakeAgent:
|
||||
def __init__(self, *args, **kwargs):
|
||||
self.valid_tool_names = {"read_file", "terminal"}
|
||||
self.session_input_tokens = 0
|
||||
self.session_output_tokens = 0
|
||||
self.session_cache_read_tokens = 0
|
||||
self.session_cache_write_tokens = 0
|
||||
self.session_api_calls = 0
|
||||
self.base_url = ""
|
||||
self.api_key = None
|
||||
|
||||
def run_conversation(self, *args, **kwargs):
|
||||
raise AssertionError("run_conversation should not be called for unavailable tools")
|
||||
|
||||
tc = ToolCall(
|
||||
id="mcp-01",
|
||||
category="mcp",
|
||||
prompt="Use an MCP tool to list resources.",
|
||||
expected_tool="",
|
||||
expected_tool_prefix="mcp_",
|
||||
)
|
||||
|
||||
with patch.dict(sys.modules, {"run_agent": SimpleNamespace(AIAgent=FakeAgent)}):
|
||||
result = run_single_test(tc, "gemini:gemma-4-31b-it", "gemini")
|
||||
|
||||
assert result.skipped is True
|
||||
assert "mcp_" in result.skip_reason
|
||||
assert result.success is False
|
||||
Reference in New Issue
Block a user