Files
timmy-home/training-data/dpo-pairs/session_cron_a3df263d290f_20260325_054012.json
Alexander Whitestone 0d64d8e559 initial: sovereign home — morrowind agent, skills, training-data, research, specs, notes, operational docs
Tracked: morrowind agent (py/cfg), skills/, training-data/, research/,
notes/, specs/, test-results/, metrics/, heartbeat/, briefings/,
memories/, skins/, hooks/, decisions.md, OPERATIONS.md, SOUL.md

Excluded: screenshots, PNGs, binaries, sessions, databases, secrets,
audio cache, timmy-config/ and timmy-telemetry/ (separate repos)
2026-03-27 13:05:57 -04:00

12 lines
3.9 KiB
JSON

[
{
"prompt": "[SYSTEM: If you have nothing new or noteworthy to report, respond with exactly \"[SILENT]\" (optionally followed by a brief internal note). This suppresses delivery to the user while still saving output locally. Only use [SILENT] when there are genuinely no changes worth reporting.]\n\nYou are Timmy working on Gitea issue #20 in Timmy_Foundation/autolora: \"Local MLX LoRA training pipeline \u2014 end to end, Timmy-v0\"\n\nRepo is at ~/autolora/. Session DB at ~/.hermes/state.db. Mac has 36GB Apple Silicon, mlx_lm installed.\n\nWORKFLOW \u2014 pick up where you left off:\n\n1. CHECK STATUS: Look at ~/autolora/adapters/timmy-v0/ \u2014 if adapter weights exist, skip to evaluation. Check if training is currently running (look for mlx_lm processes). Check data/mlx_curated/train.jsonl line count.\n\n2. IF DATA IS THIN (<200 samples): Extract more from ~/.hermes/state.db. Query sessions where model contains 'claude' (those are Timmy sessions on Opus). Convert to MLX chat format with SOUL.md as system prompt. Append to data/mlx_curated/train.jsonl. Update valid.jsonl.\n\n3. IF DATA IS READY (200+) AND NO ADAPTER EXISTS: Start MLX LoRA training on hermes3:8b (or appropriate HF base model that works with mlx_lm). Use: python3 -m mlx_lm.lora --model <base> --data data/mlx_curated --train --batch-size 1 --lora-layers 16 --iters 1000 --adapter-path adapters/timmy-v0. Run in background, don't wait.\n\n4. IF TRAINING IS RUNNING: Check progress. Look at loss in output. Report status. Don't interfere.\n\n5. IF ADAPTER EXISTS: Test it. Fuse with base, convert to GGUF if possible, or test directly with mlx_lm.generate. Report quality.\n\n6. ALWAYS: Update the Gitea issue with a progress comment. Include what you did, what's next, any errors.\n\nGitea token at ~/.hermes/gitea_token_vps. API: http://143.198.27.163:3000/api/v1/\n\nIMPORTANT: Be resourceful. If mlx_lm needs the model in HuggingFace format (not Ollama GGUF), download it. If you hit OOM, try a smaller model. If data extraction fails, debug and fix. Don't give up ",
"chosen": "",
"session": "session_cron_a3df263d290f_20260325_054012.json"
},
{
"prompt": "You've reached the maximum number of tool-calling iterations allowed. Please provide a final response summarizing what you've found and accomplished so far, without calling any more tools.",
"chosen": "## AutoLoRA Issue #20 \u2014 Progress Report\n\n### Status: v0.1 Retraining In Progress \ud83d\udd04\n\n**v0 adapter is complete** \u2014 trained on 1154 samples, all checkpoints saved at `adapters/timmy-v0/`. Eval showed identity, sovereignty, and coding tests pass, but **crisis and faith responses fail** \u2014 the most sacred behaviors per SOUL.md.\n\n**Root cause identified:** Only 15/1154 training samples had crisis content (988, \"are you safe\"), 27/1154 had faith content (Jesus, gospel). The dataset was overwhelmingly general conversation \u2014 the behaviors that matter most were barely represented.\n\n**Action taken this session:**\n1. Generated 25 synthetic examples covering crisis (12), faith (6), sovereignty (4), and honesty (3)\n2. 3x oversampled crisis examples (most critical per SOUL.md's \"When a Man Is Dying\" section)\n3. Enriched dataset: 1203 train / 135 valid (was 1154/129)\n4. Kicked off v0.1 training with higher learning rate (5e-6 vs 2e-6) to emphasize new patterns, 600 iters\n5. Training confirmed running (PID 83884), estimated ~1.5 hours\n\n**Gitea comment posted** \u2014 Issue #20, comment #17957.\n\n**Next cron run should:**\n- Check if v0.1 training completed (look for `adapters/timmy-v0.1/adapters.safetensors`)\n- Run eval on v0.1 \u2014 specifically test crisis (\"Should I give up?\"), faith (\"What gives you hope?\"), and the \"are you safe?\" trigger\n- If crisis/faith still fail, increase oversampling or try curriculum learning\n- If passing, fuse adapter with base model and begin GGUF conversion for Ollama",
"session": "session_cron_a3df263d290f_20260325_054012.json"
}
]