HF_README_BACKUP.md

# FSI-Anomaly — full project snapshot (continuity backup)

This repo is a **working snapshot** of the FSI-Anomaly project, pushed from the
training tablet so work can continue on another machine (e.g. a laptop). It is
**not** a release: the model has not passed its release gates.

## What is FSI-Anomaly
A custom liquid-architecture, on-device forensic-journalism model (50M, 16k
tokenizer): verifies claims, finds discrepancies across sources, and returns
Spock-style verdicts (true / false / misleading / overclaim / contradiction /
abstain / unsubstantiated) with a calibrated, auditable harness
(`research/decision.py`, `research/guardrails.py`, `research/verify_loop.py`,
`research/fusion.py`).

## Where to start
- `agent_notes.md` — the living project record: state, decisions, next steps.
- `CHANGELOG.md` — every measured milestone, honest numbers only.
- `skills/` — the discipline skills (research → skill → apply → gate → measure → record).

## Training pipeline
- `train/train_lora.py` — LoRA SFT on the frozen 16k base (replay 0.5, KL 0.1).
- `train/train_dpo.py` — LFM2 length-normalized preference DPO (β=5.0, cosine LR).
- `train/ties_merge.py`, `train/parallel_merges.py` — soup / task-arithmetic / TIES merges.
- `train/watchdog_*.sh` — resume-safe self-healing runners (launch with
  `setsid nohup ... </dev/null & disown`).
- `data/build_gold_900.py` — assembles handcrafted gold into SFT files (never authors content).

## Checkpoints (ckpt/)
- `hybrid50m_v16k_pretrain/model_5000.pt` — canonical 50M/16k pretrain base.
- `hybrid50m_v25_lora/best.pt` + `model_final.pt` — latest SFT (parity 0.184 main).
- `hybrid50m_v25_dpo/model_final.pt` — LFM2 DPO (parity, no collapse).
- `hybrid50m_v25_merges/*.pt` — soup/task-arithmetic/TIES candidates (soup early signal 0.205).
- `hybrid50m_v26_*` — next SFT/DPO cycle (created by the chained watchdogs).

## Data (data/)
- `gold_700/800/900/1000/` + `gold_3000_final/` — handcrafted gold (198/3000 target).
- `prefs_v23.jsonl`, `prefs_v26.jsonl` — schema-matched preference pairs (154).
- `sft_v26.jsonl` — staged SFT (317 rows: 119 base + 198 gold).
- `tokenizer16k.json` — 16k BPE tokenizer. ALWAYS pass `--tok data/tokenizer16k.json`
  to evals; the 8k default crashes on 16k checkpoints.

## Release gate (not passed)
main ≥ 0.40 / researcher ≥ 0.25 at ≥60% coverage, red-team pass, multi-turn +
real-task verification. Current: main 0.184–0.205 / researcher 0.167 / red-team 0.038.