docs/research/2026-07-22-vtd-handoff.md

VTD tiny-model handoff — 2026-07-22 (PC crashed under load; state pushed)

Branch: claude/spiral-harness (PR #2832 tail; the main body merged in #2823). Everything below is committed here. The PC was crashing under combined load (Ollama + exec subprocesses + installs), so all my background jobs are dead/killed; nothing is left running.

Where the science stands (all exec-verified; ledger records cited)

experiment dose config held-out MBPP [400–450) verdict
baseline 0.5B (Qwen2.5-Coder-0.5B-Instruct) — — 21/50 (42%) fragile but real
retrieval (few-shot own traces) 6 traces — 6/6 → 2/6 on DP set harmful (template contamination)
VTD run(cr-mrvvxsuc) 63 traces lr 2e-4, 6ep, r16 15/50, −6 harmful (memorized, damaged instruct)
VTD run(cr-mrvxt1li) 204 traces lr 5e-5, 3ep, r8 21/50, ±0 (3 fixed /regressed) neutral — dose-response confirmed

Direction: more data + gentler = less damage + real fixes. Crossover to a net lift needs the next order of magnitude of traces → TACO.

Artifacts on this branch

  • data/eval/spiral/vtd-corpus-all.jsonl —verified MBPP traces (63 + 141;frontier rescues).
  • data/eval/spiral/vtd-taco-1.jsonl — 30 partial TACO traces (batchdied at ~34/400 in the crash).
  • data/eval/taco-easy.jsonl — 1,581 normalized TACO-verified EASY stdio problems (via

scripts/fetch_taco.py; source likaixin/TACO-verified, 12,898 rows, canonical BAAI/TACO's loading script is unsupported by modern datasets).

  • data/eval/mbpp-full.jsonl —normalized MBPP problems (scripts/fetch_mbpp.py).

Held-out slices — never train on: MBPP [400–450); reserve TACO [1400–1500) similarly.

Artifacts on disk (NOT in git)

  • Adapters: D:\lantern-train\qwen05-vtd\final (run 1), D:\lantern-train\qwen05-vtd2\final (run 2).
  • Base model cached: D:\hf-cache (Qwen/Qwen2.5-Coder-0.5B-Instruct).
  • Logs: D:\lantern-train\*-train.log, *-eval.log, gen*.log.

⚠ Gotchas the next session must know

  1. LlamaFactory venv (D:\venvs\llamafactory) has a BROKEN torch: pip's resolver replaced the

cu121 torch with 2.13.0+cpu (cuda False) during pip install llamafactory. Fix before use: D:\venvs\llamafactory\Scripts\pip install torch==2.5.1 --index-url https://download.pytorch.org/whl/cu121 --force-reinstall then verify torch.cuda.is_available(). (LF 0.9.5 + transformers 5.6.0 installed.)

  1. One heavy job at a time on this box. The crash came from concurrent load. Unload Ollama models

first (POST /api/generate {"model":..., "keep_alive":0}), run trainers/generators detached (PowerShell Start-Process) — an attached run died mid-CUDA once already ("device not ready").

  1. Licensing (founder policy, cr-mrwcmob0): unisona is not commercial yet — KodCode data

(CC BY-NC) is research-OK but tag any weights trained on it nc-contaminated; TACO (Apache) is the default corpus for product-lineage weights.

  1. The tier-preamble stdio bug is FIXED (c6cc6fba) — cheap sufficiency on TACO went 40%→80%.

Escalation is now honest (cloud→cloud fallback only; fails rather than degrading to local).

Resume plan (run 3)

  1. Finish TACO batch in small chunks (crash-safe), e.g.:

SPIRAL_FRONTIER_PROVIDER=openai node experiments/spiral_gen_traces.js --src data/eval/taco-easy.jsonl --limit--offset--out data/eval/spiral/vtd-taco-1b.jsonl (repeat offsets 134, 234, 334; merge with vtd-corpus-all.jsonl + vtd-taco-1.jsonl).

  1. Fix the LF venv torch (above). Convert merged corpus → LlamaFactory alpaca format + a **retention

mix** (~400 rows of a permissive general-instruct set, e.g. dolly-15k) via dataset_info.json; train QLoRA r=8, lr 5e-5, ≤3 epochs, qwen template, val_size + best-checkpoint.

  1. Eval with scripts/eval_qwen_coder.py on MBPP [400–450) (comparable to −6 → ±0) AND a TACO

held-out slice; emit the ConvergenceRecord either way (honest null included).