Skip to content

methodology.md

Latest commit

 

History

History
57 lines (44 loc) · 2.59 KB

methodology.md

File metadata and controls

57 lines (44 loc) · 2.59 KB
Raw

Suture — Methodology (shared across the four forge repos)

This is the recipe that already beat a frontier model with a 1.7B LoRA (ReconForge, ~/Documents/personal/reconforge) — now scaled to a 28B multimodal Qwen on cloud free credits.

1. Seeded synthetic data + verifier-as-oracle

  • A seeded generator draws true facts, injects a known violation set, and derives "as printed" fields that realize exactly that set.
  • A deterministic oracle recomputes the violations from the printed fields — no model, no LLM judge.
  • A self-check gate runs every candidate task through the oracle and regenerates on disagreement, so oracle agreement is a property.
  • Same seed → same tasks (byte-identical signatures).

2. Contamination control

  • Task signature = SHA-256 over sorted ground-truth fields.
  • Train/val/benchmark are signature-disjoint; a leak probe must fire 1.0 on intentional leaks and 0.0 on clean sets (ROC).

3. Training (SOTA 2026)

  1. SFT cold-start: QLoRA 4-bit on verifier-filtered traces distilled from a stronger model. RLVR data never mixed in.
  2. RLVR: GRPO with the Dr. GRPO fix (drop per-token length normalization), DAPO-style decoupled clip (high 1.0 / low 0.2), dynamic sampling of non-saturated prompts, G≈8–16, outcome-verifier-as-oracle only (no PRM — narrow, machine-verifiable output; safest against reward hacking).
  3. Self-play: ReST-EM / expert iteration rounds + one s1-style curation pass (~1k high-difficulty traces).
  4. Judges: pairwise, position-swapped, temp 0, majority of 3, bootstrap CIs, kappa ≥ 0.85 vs a golden-100 set. Only for residual prose.

4. Evaluation

  • Core metrics are deterministic (oracle-scored). Head-to-head vs frontier (Qwen3.8-2.4T-A95B, DeepSeek v4-flash) and vs the base model zero-shot, on identical inputs and scoring code.
  • Frontier eval runs concurrent with JSONL checkpointing (resumable).

5. Cloud (free credits)

  • Modal primary: your-modal-profile profile, $30/mo recurring credit, L4 24GB ≈ 37 GPU-hrs/mo, free 1 TiB volume for resumable checkpoints.
  • GCP fallback: your-gcp-project (you@example.com), Compute API enabled, GPU quota limit 1000. Spot L4/A100 marathon.
  • Gate before training: smoke LoRA on the base model to validate the 3-day-old Qwen3.8-27B DeltaNet fine-tuning tooling; fallback TRL+PEFT or the FP8 base.

6. Honest reporting

Every repo records negative results in docs/DECISIONS.md with evidence. Base-model and eval numbers are self-measured on self-built benchmarks; that is the point — the methodology is published with the numbers.