[AutoLoRA P1] Load Fine-Tuned Model Back Into Hermes Harness #1104

Closed
opened 2026-03-23 17:29:44 +00:00 by perplexity · 1 comment
Collaborator

Load Fine-Tuned Model Back Into Hermes Harness

Priority: P1-Important
Assignee: Alexander
Epic: #1091 — Project Bannerlord
Pipeline: AutoLoRA Sovereignty Loop (Step 5 of 7)
Blocked by: LoRA Fine-Tune issue


What To Do

1. Merge the LoRA adapter into the base model:

mlx_lm.fuse \
  --model <path-to-hermes4-model> \
  --adapter-path ~/timmy-lora-adapter \
  --save-path ~/timmy-fused-model

2. Convert to GGUF for Ollama (if fuse doesn't output GGUF directly):

python3 llama.cpp/convert_hf_to_gguf.py ~/timmy-fused-model --outtype q5_k_m

3. Import into Ollama:

cat > Modelfile.timmy << 'EOF'
FROM ./timmy-fused-model.Q5_K_M.gguf
PARAMETER num_ctx 32768
PARAMETER temperature 0.3
SYSTEM "You are Timmy, a sovereign AI agent. You operate inside the Hermes Agent harness."
EOF

ollama create timmy -f Modelfile.timmy

4. Switch the harness:

hermes model timmy

5. Test every skill. Document what works and what still fails.

Done When

  • ollama list shows a model named timmy
  • hermes model timmy switches successfully
  • At least 25 of 32 skills work correctly
  • Skills that still fail are documented as separate issues
## Load Fine-Tuned Model Back Into Hermes Harness **Priority:** P1-Important **Assignee:** Alexander **Epic:** #1091 — Project Bannerlord **Pipeline:** AutoLoRA Sovereignty Loop (Step 5 of 7) **Blocked by:** LoRA Fine-Tune issue --- ### What To Do **1. Merge the LoRA adapter into the base model:** ```bash mlx_lm.fuse \ --model <path-to-hermes4-model> \ --adapter-path ~/timmy-lora-adapter \ --save-path ~/timmy-fused-model ``` **2. Convert to GGUF for Ollama (if fuse doesn't output GGUF directly):** ```bash python3 llama.cpp/convert_hf_to_gguf.py ~/timmy-fused-model --outtype q5_k_m ``` **3. Import into Ollama:** ```bash cat > Modelfile.timmy << 'EOF' FROM ./timmy-fused-model.Q5_K_M.gguf PARAMETER num_ctx 32768 PARAMETER temperature 0.3 SYSTEM "You are Timmy, a sovereign AI agent. You operate inside the Hermes Agent harness." EOF ollama create timmy -f Modelfile.timmy ``` **4. Switch the harness:** ```bash hermes model timmy ``` **5. Test every skill. Document what works and what still fails.** ### Done When - [ ] `ollama list` shows a model named `timmy` - [ ] `hermes model timmy` switches successfully - [ ] At least 25 of 32 skills work correctly - [ ] Skills that still fail are documented as separate issues
claude self-assigned this 2026-03-23 17:54:04 +00:00
Collaborator

PR created: http://143.198.27.163:3000/rockachopa/Timmy-time-dashboard/pulls/1122

Changes:

  • Modelfile.timmy — Ollama Modelfile for the LoRA-fused model (Hermes 4 14B + Timmy LoRA adapter, Q5_K_M, 32K ctx, Timmy persona system prompt)
  • scripts/fuse_and_load.sh — Pipeline automation: mlx_lm.fusellama.cpp GGUF conversion → ollama create timmy. Env var overrides for all paths, SKIP_FUSE/SKIP_CONVERT flags.
  • scripts/test_timmy_skills.py — 32-skill validation suite (threshold 25/32). Covers all tool-call skills, JSON, reasoning, code gen, multi-turn, safety, persona identity.
  • config/providers.yamltimmy model registered and placed at the top of the tools fallback chain.

Run order:

  1. ./scripts/fuse_and_load.sh — fuse + import
  2. hermes model timmy — switch harness
  3. python scripts/test_timmy_skills.py — validate 32 skills, file issues for failures
PR created: http://143.198.27.163:3000/rockachopa/Timmy-time-dashboard/pulls/1122 **Changes:** - `Modelfile.timmy` — Ollama Modelfile for the LoRA-fused model (Hermes 4 14B + Timmy LoRA adapter, Q5_K_M, 32K ctx, Timmy persona system prompt) - `scripts/fuse_and_load.sh` — Pipeline automation: `mlx_lm.fuse` → `llama.cpp` GGUF conversion → `ollama create timmy`. Env var overrides for all paths, `SKIP_FUSE`/`SKIP_CONVERT` flags. - `scripts/test_timmy_skills.py` — 32-skill validation suite (threshold 25/32). Covers all tool-call skills, JSON, reasoning, code gen, multi-turn, safety, persona identity. - `config/providers.yaml` — `timmy` model registered and placed at the top of the `tools` fallback chain. **Run order:** 1. `./scripts/fuse_and_load.sh` — fuse + import 2. `hermes model timmy` — switch harness 3. `python scripts/test_timmy_skills.py` — validate 32 skills, file issues for failures
Sign in to join this conversation.
No Label
2 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: Rockachopa/Timmy-time-dashboard#1104