feat: add training data quality filter (#687)

bin/nostr-agent-demo.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 """
 Full Nostr agent-to-agent communication demo - FINAL WORKING
 """

bin/soul_eval_gate.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 """
 Soul Eval Gate — The Conscience of the Training Pipeline
 

scripts/captcha_bypass_handler.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import json
 from hermes_tools import browser_navigate, browser_vision
 

scripts/diagram_meaning_extractor.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import json
 from hermes_tools import browser_navigate, browser_vision
 

scripts/pr_triage.py

@@ -1,176 +0,0 @@
-#!/usr/bin/env python3
-"""PR Triage Automation -- Categorize, deduplicate, report (#659)."""
-import argparse, json, os, re, sys, subprocess
-from collections import Counter, defaultdict
-from datetime import datetime
-from urllib.request import Request, urlopen
-from urllib.error import HTTPError
-
-
-def _token():
-    t = os.environ.get("GITEA_TOKEN", "")
-    if not t:
-        p = os.path.expanduser("~/.config/gitea/token")
-        if os.path.exists(p):
-            t = open(p).read().strip()
-    return t
-
-
-def _api(url, token, method="GET", data=None):
-    h = {"Authorization": "token " + token, "Accept": "application/json"}
-    body = json.dumps(data).encode() if data else None
-    if data:
-        h["Content-Type"] = "application/json"
-    req = Request(url, data=body, headers=h, method=method)
-    try:
-        return json.loads(urlopen(req, timeout=30).read())
-    except HTTPError:
-        return None
-
-
-def fetch_prs(base, token, owner, repo):
-    prs, page = [], 1
-    while True:
-        b = _api(base + "/api/v1/repos/" + owner + "/" + repo + "/pulls?state=open&limit=50&page=" + str(page), token)
-        if not b:
-            break
-        prs.extend(b)
-        if len(b) < 50:
-            break
-        page += 1
-    return prs
-
-
-def fetch_issues(base, token, owner, repo):
-    iss, page = {}, 1
-    while True:
-        b = _api(base + "/api/v1/repos/" + owner + "/" + repo + "/issues?state=open&limit=50&page=" + str(page), token)
-        if not b:
-            break
-        for i in b:
-            if "pull_request" not in i:
-                iss[i["number"]] = i
-        if len(b) < 50:
-            break
-        page += 1
-    return iss
-
-
-def categorize(pr):
-    c = (pr.get("title", "") + " " + pr.get("body", "") + " " + " ".join(l.get("name", "") for l in pr.get("labels", []))).lower()
-    for kw, cat in [("training data", "training-data"), ("dpo", "training-data"), ("grpo", "training-data"),
-                    ("fix:", "bug-fix"), ("bug", "bug-fix"), ("hotfix", "bug-fix"),
-                    ("feat:", "feature"), ("feature", "feature"),
-                    ("refactor", "maintenance"), ("cleanup", "maintenance"),
-                    ("doc", "documentation"), ("test", "testing"), ("infra", "infrastructure")]:
-        if kw in c:
-            return cat
-    return "other"
-
-
-def refs(pr):
-    return [int(m) for m in re.findall(r"#(\d+)", pr.get("title", "") + " " + pr.get("body", ""))]
-
-
-def find_duplicates(prs):
-    by = defaultdict(list)
-    for p in prs:
-        for r in refs(p):
-            by[r].append(p)
-    return [g for g in by.values() if len(g) > 1]
-
-
-def health(pr, issues):
-    r = refs(pr)
-    created = datetime.fromisoformat(pr["created_at"].replace("Z", "+00:00"))
-    updated = datetime.fromisoformat(pr["updated_at"].replace("Z", "+00:00"))
-    now = datetime.now(created.tzinfo)
-    return {
-        "pr": pr["number"], "title": pr["title"], "head": pr["head"]["ref"],
-        "category": categorize(pr), "refs": r,
-        "open": [x for x in r if x in issues], "closed": [x for x in r if x not in issues],
-        "age": (now - created).days, "stale": (now - updated).days,
-        "mergeable": pr.get("mergeable"), "author": pr.get("user", {}).get("login", ""),
-    }
-
-
-def report(repo, checks, dups):
-    lines = ["# PR Triage -- " + repo,
-             "Generated: " + datetime.now().strftime("%Y-%m-%d %H:%M"),
-             "Open PRs: " + str(len(checks)), "", "## Summary", ""]
-    cats = Counter(h["category"] for h in checks)
-    lines.append("| Category | Count |")
-    lines.append("|----------|-------|")
-    for c, n in cats.most_common():
-        lines.append("| " + c + " | " + str(n) + " |")
-    stale = [h for h in checks if h["stale"] > 7]
-    lines.extend(["", "Stale (>7d): " + str(len(stale)),
-                  "Duplicate groups: " + str(len(dups)), ""])
-    if dups:
-        lines.append("## Duplicates")
-        for g in dups:
-            rs = set()
-            for p in g:
-                rs.update(refs(p))
-            lines.append("Issues " + ", ".join("#" + str(r) for r in sorted(rs)) + ":")
-            for p in g:
-                lines.append("  - #" + str(p["number"]) + ": " + p["title"])
-        lines.append("")
-    if stale:
-        lines.append("## Stale (>7d)")
-        for h in sorted(stale, key=lambda x: x["stale"], reverse=True):
-            lines.append("- #" + str(h["pr"]) + ": " + h["title"] + " -- " + str(h["stale"]) + "d")
-        lines.append("")
-    lines.append("## All PRs")
-    lines.append("| # | Title | Category | Age | Stale | Merge |")
-    lines.append("|---|-------|----------|-----|-------|-------|")
-    for h in sorted(checks, key=lambda x: x["pr"]):
-        m = "Y" if h["mergeable"] else ("N" if h["mergeable"] is False else "?")
-        s = str(h["stale"]) + "d" if h["stale"] > 7 else "-"
-        lines.append("| " + str(h["pr"]) + " | " + h["title"][:50] + " | " + h["category"] +
-                      " | " + str(h["age"]) + "d | " + s + " | " + m + " |")
-    return chr(10).join(lines)
-
-
-def main():
-    p = argparse.ArgumentParser(description="PR Triage Automation")
-    p.add_argument("--base-url", default="https://forge.alexanderwhitestone.com")
-    p.add_argument("--owner", default="Timmy_Foundation")
-    p.add_argument("--repo", default="")
-    p.add_argument("--json", action="store_true", dest="js")
-    p.add_argument("--output", default="")
-    a = p.parse_args()
-    token = _token()
-    if not token:
-        print("No token"); sys.exit(1)
-    repo = a.repo
-    if not repo:
-        try:
-            remote = subprocess.check_output(["git", "remote", "get-url", "origin"], text=True).strip()
-            m = re.search(r"[/:](\w[\w-]*)/(\w[\w-]*?)(?:\.git)?$", remote)
-            if m:
-                a.owner, repo = m.group(1), m.group(2)
-        except Exception:
-            pass
-    if not repo:
-        print("No repo specified"); sys.exit(1)
-    print("Triaging " + a.owner + "/" + repo + "...", file=sys.stderr)
-    prs = fetch_prs(a.base_url, token, a.owner, repo)
-    issues = fetch_issues(a.base_url, token, a.owner, repo)
-    checks = [health(pr, issues) for pr in prs]
-    dups = find_duplicates(prs)
-    if a.js:
-        print(json.dumps({"repo": repo, "prs": checks,
-                          "duplicates": [[{"number": p["number"], "title": p["title"]} for p in g] for g in dups]},
-                         indent=2))
-    else:
-        r = report(repo, checks, dups)
-        print(r)
-        if a.output:
-            with open(a.output, "w") as f:
-                f.write(r)
-    print("\n" + str(len(checks)) + " PRs, " + str(len(dups)) + " duplicate groups", file=sys.stderr)
-
-
-if __name__ == "__main__":
-    main()

scripts/training_data_quality_filter.py

@@ -0,0 +1,286 @@
+#!/usr/bin/env python3
+"""
+Training Data Quality Filter
+
+Scores and removes low-quality training pairs from JSONL datasets.
+Supports two formats:
+  - ShareGPT session format: {"conversations": [...], ...}
+  - Scene/pair format: {"terse": "...", "rich": "..."} or {"lyric_line": "...", "scene": {...}}
+
+Scoring dimensions:
+  - Specificity: penalizes vague/generic content
+  - Length ratio: penalizes extreme input/output imbalances
+  - Code correctness: validates code blocks have matching fences
+
+Usage:
+  python3 scripts/training_data_quality_filter.py input.jsonl [--threshold 0.4] [--output filtered.jsonl]
+  python3 scripts/training_data_quality_filter.py --dir training-data/ [--threshold 0.4]
+"""
+
+import argparse
+import json
+import re
+import sys
+from pathlib import Path
+
+
+def score_specificity(text: str) -> float:
+    """Score 0-1 based on how specific vs generic the text is."""
+    if not text or len(text.strip()) < 10:
+        return 0.0
+
+    score = 0.5  # baseline
+
+    # Penalize very generic starters
+    generic_starters = [
+        "sure,", "of course", "i can help", "here is", "here are",
+        "certainly", "absolutely", "let me help", "great question",
+        "that\'s a great", "interesting question",
+    ]
+    lower = text.lower().strip()
+    for starter in generic_starters:
+        if lower.startswith(starter):
+            score -= 0.15
+            break
+
+    # Reward specific content indicators
+    if re.search(r"`[^`]+`", text):  # inline code
+        score += 0.1
+    if re.search(r"```[\s\S]*?```", text):  # code blocks
+        score += 0.15
+    if re.search(r"\d+\.\s", text):  # numbered lists
+        score += 0.05
+    if len(text.split()) > 50:  # substantial length
+        score += 0.1
+    if re.search(r"https?://", text):  # URLs/references
+        score += 0.05
+
+    # Penalize extremely short outputs
+    if len(text.split()) < 5:
+        score -= 0.2
+
+    # Penalize repetition (same sentence repeated)
+    sentences = re.split(r"[.!?]+", text)
+    sentences = [s.strip().lower() for s in sentences if s.strip()]
+    if sentences:
+        unique_ratio = len(set(sentences)) / len(sentences)
+        if unique_ratio < 0.7:
+            score -= 0.15
+
+    return max(0.0, min(1.0, score))
+
+
+def score_length_ratio(input_text: str, output_text: str) -> float:
+    """Score 0-1 based on input/output length balance."""
+    in_len = len(input_text.split())
+    out_len = len(output_text.split())
+
+    if in_len == 0 or out_len == 0:
+        return 0.0
+
+    ratio = out_len / in_len
+
+    # Ideal ratio: 0.5-5x (output can be shorter or longer, but not extreme)
+    if 0.5 <= ratio <= 5.0:
+        return 1.0
+    elif 0.2 <= ratio <= 10.0:
+        return 0.6
+    elif 0.1 <= ratio <= 20.0:
+        return 0.3
+    else:
+        return 0.1
+
+
+def score_code_correctness(text: str) -> float:
+    """Score 0-1 based on code block correctness."""
+    code_blocks = re.findall(r"```[\s\S]*?```", text)
+
+    if not code_blocks:
+        return 1.0  # no code = no code errors
+
+    for block in code_blocks:
+        # Check balanced fences
+        fence_count = block.count("```")
+        if fence_count % 2 != 0:
+            return 0.2
+
+        # Check for common errors
+        content = block.split("\n", 1)[-1] if "\n" in block else ""
+        if "SyntaxError" in content or "Traceback" in content:
+            return 0.3
+        if content.strip().endswith("...") and len(content.strip()) < 30:
+            return 0.4  # truncated code
+
+    return 1.0
+
+
+def score_pair(input_text: str, output_text: str) -> dict:
+    """Score a training pair on all dimensions."""
+    spec = score_specificity(output_text)
+    length = score_length_ratio(input_text, output_text)
+    code = score_code_correctness(output_text)
+
+    # Weighted composite
+    composite = (spec * 0.4) + (length * 0.3) + (code * 0.3)
+
+    return {
+        "specificity": round(spec, 3),
+        "length_ratio": round(length, 3),
+        "code_correctness": round(code, 3),
+        "composite": round(composite, 3),
+    }
+
+
+def extract_pairs(obj: dict) -> list:
+    """Extract (input, output) pairs from a JSONL object."""
+    pairs = []
+
+    # ShareGPT session format
+    if "conversations" in obj:
+        convs = obj["conversations"]
+        for i, msg in enumerate(convs):
+            if msg.get("from") in ("gpt", "assistant"):
+                # Find preceding human message
+                input_text = ""
+                for j in range(i - 1, -1, -1):
+                    if convs[j].get("from") == "human":
+                        input_text = convs[j].get("value", "")
+                        break
+                output_text = msg.get("value", "")
+                if input_text and output_text:
+                    pairs.append((input_text, output_text))
+
+    # Scene/pair format (terse/rich)
+    elif "terse" in obj and "rich" in obj:
+        pairs.append((obj["terse"], obj["rich"]))
+
+    # Scene description format
+    elif "lyric_line" in obj and "scene" in obj:
+        scene_text = json.dumps(obj["scene"]) if isinstance(obj["scene"], dict) else str(obj["scene"])
+        pairs.append((obj["lyric_line"], scene_text))
+
+    # Generic prompt/response
+    elif "prompt" in obj and "response" in obj:
+        pairs.append((obj["prompt"], obj["response"]))
+
+    # Generic input/output
+    elif "input" in obj and "output" in obj:
+        pairs.append((obj["input"], obj["output"]))
+
+    return pairs
+
+
+def filter_jsonl(input_path: str, threshold: float = 0.4, output_path: str = None) -> dict:
+    """Filter a JSONL file, removing low-quality pairs."""
+    path = Path(input_path)
+    if not path.exists():
+        return {"error": f"File not found: {input_path}"}
+
+    lines = path.read_text().strip().split("\n")
+    total = 0
+    kept = 0
+    removed = 0
+    scores_list = []
+    kept_lines = []
+
+    for line in lines:
+        line = line.strip()
+        if not line:
+            continue
+
+        try:
+            obj = json.loads(line)
+        except json.JSONDecodeError:
+            removed += 1
+            continue
+
+        pairs = extract_pairs(obj)
+        total += 1
+
+        if not pairs:
+            # No extractable pairs — keep as-is (might be metadata)
+            kept += 1
+            kept_lines.append(line)
+            continue
+
+        # Score all pairs in this object
+        pair_scores = [score_pair(inp, out) for inp, out in pairs]
+        avg_composite = sum(s["composite"] for s in pair_scores) / len(pair_scores)
+
+        scores_list.append(avg_composite)
+
+        if avg_composite >= threshold:
+            kept += 1
+            kept_lines.append(line)
+        else:
+            removed += 1
+
+    # Write output
+    if output_path:
+        Path(output_path).write_text("\n".join(kept_lines) + "\n")
+
+    return {
+        "file": input_path,
+        "total": total,
+        "kept": kept,
+        "removed": removed,
+        "removal_rate": f"{removed}/{total}" if total > 0 else "0/0",
+        "avg_score": round(sum(scores_list) / len(scores_list), 3) if scores_list else None,
+        "min_score": round(min(scores_list), 3) if scores_list else None,
+        "max_score": round(max(scores_list), 3) if scores_list else None,
+    }
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Filter low-quality training data pairs")
+    parser.add_argument("input", nargs="?", help="Input JSONL file")
+    parser.add_argument("--threshold", type=float, default=0.4, help="Minimum quality score (0-1)")
+    parser.add_argument("--output", "-o", help="Output file (default: input_filtered.jsonl)")
+    parser.add_argument("--dir", help="Process all .jsonl files in directory")
+    parser.add_argument("--dry-run", action="store_true", help="Score only, don\'t write output")
+    args = parser.parse_args()
+
+    if args.dir:
+        dirpath = Path(args.dir)
+        jsonl_files = sorted(dirpath.rglob("*.jsonl"))
+        if not jsonl_files:
+            print(f"No .jsonl files found in {args.dir}")
+            sys.exit(1)
+
+        print(f"Processing {len(jsonl_files)} files (threshold={args.threshold})\n")
+        print(f"{'File':<50} {'Total':>6} {'Kept':>6} {'Removed':>8} {'Avg':>6}")
+        print("-" * 82)
+
+        grand_total = grand_kept = grand_removed = 0
+        for fpath in jsonl_files:
+            out = str(fpath).replace(".jsonl", "_filtered.jsonl") if not args.dry_run else None
+            result = filter_jsonl(str(fpath), args.threshold, out)
+            if "error" in result:
+                print(f"{str(fpath):<50} ERROR: {result['error']}")
+                continue
+            print(f"{fpath.name:<50} {result['total']:>6} {result['kept']:>6} {result['removed']:>8} {result['avg_score']:>6.3f}")
+            grand_total += result["total"]
+            grand_kept += result["kept"]
+            grand_removed += result["removed"]
+
+        print("-" * 82)
+        print(f"{'TOTAL':<50} {grand_total:>6} {grand_kept:>6} {grand_removed:>8}")
+
+    elif args.input:
+        out = args.output or args.input.replace(".jsonl", "_filtered.jsonl")
+        if args.dry_run:
+            out = None
+        result = filter_jsonl(args.input, args.threshold, out)
+        if "error" in result:
+            print(f"Error: {result['error']}")
+            sys.exit(1)
+        print(json.dumps(result, indent=2))
+        if out:
+            print(f"\nFiltered output written to: {out}")
+    else:
+        parser.print_help()
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()

scripts/visual_pr_reviewer.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import json
 from hermes_tools import browser_navigate, browser_vision
 

tests/test_pr_triage.py

@@ -1,45 +0,0 @@
-"""Tests for PR triage automation (#659)."""
-import pytest
-
-class TestCategorize:
-    def _pr(self, title="", body=""):
-        return {"title": title, "body": body, "labels": []}
-
-    def test_training(self):
-        from scripts.pr_triage import categorize
-        assert categorize(self._pr("Add DPO pairs")) == "training-data"
-
-    def test_bug(self):
-        from scripts.pr_triage import categorize
-        assert categorize(self._pr("fix: crash")) == "bug-fix"
-
-    def test_feature(self):
-        from scripts.pr_triage import categorize
-        assert categorize(self._pr("feat: dark mode")) == "feature"
-
-    def test_other(self):
-        from scripts.pr_triage import categorize
-        assert categorize(self._pr("random")) == "other"
-
-class TestRefs:
-    def test_simple(self):
-        from scripts.pr_triage import refs
-        assert 123 in refs({"title": "Fix #123", "body": ""})
-
-    def test_multiple(self):
-        from scripts.pr_triage import refs
-        r = refs({"title": "", "body": "Closes #100, Refs #200"})
-        assert 100 in r and 200 in r
-
-class TestDuplicates:
-    def test_found(self):
-        from scripts.pr_triage import find_duplicates
-        prs = [{"title": "", "body": "Fix #1", "number": 1, "head": {"ref": "a"}, "created_at": "2026-01-01T00:00:00Z", "updated_at": "2026-01-01T00:00:00Z", "user": {}},
-               {"title": "", "body": "Refs #1", "number": 2, "head": {"ref": "b"}, "created_at": "2026-01-01T00:00:00Z", "updated_at": "2026-01-01T00:00:00Z", "user": {}}]
-        assert len(find_duplicates(prs)) == 1
-
-    def test_none(self):
-        from scripts.pr_triage import find_duplicates
-        prs = [{"title": "", "body": "Fix #1", "number": 1, "head": {"ref": "a"}, "created_at": "2026-01-01T00:00:00Z", "updated_at": "2026-01-01T00:00:00Z", "user": {}},
-               {"title": "", "body": "Fix #2", "number": 2, "head": {"ref": "b"}, "created_at": "2026-01-01T00:00:00Z", "updated_at": "2026-01-01T00:00:00Z", "user": {}}]
-        assert find_duplicates(prs) == []