Skip to content

v0.7.0 — Phase 5: Small-model amplification

Choose a tag to compare

View all tags
@hackspaces hackspaces released this 08 Jul 15:25
v0.7.0
4b21227

Phase 5 complete — the harness scaffolding that makes small local models reliable.

Highlights

  • Model passports (P5.8) — forge measures each model (active ~90s setup probe + passive live telemetry) and auto-tunes its own knobs: tighter loop threshold for loop-prone models, bigger num_predict for write-file truncators, hotter retry schedule for greedy-identical failers. New forge passport.
  • Step-scoped borrowing + unified stuck ledger (P5.7) — one weighted per-turn stuck score; buy a single strong-rung generation at the exact decision points the cheap model is stuck, then decay back.
  • Self-harvested few-shot exemplars (P5.6) — malformed-retry nudges quote the model's own past valid actions.
  • Retry-heat temperature (P5.5) — greedy first try, perturb on each nudge, reset on success.
  • Deterministic JSON salvage (P5.4) — recover fenced / prose-prefixed / trailing-comma action JSON before counting a malformed strike.
  • Line-anchored edit dialect (P5.3), dry-run verifier + best-of-N resample (P5.2), state-dependent action grammar (P5.1).

Stdlib-only, zero runtime deps. 403 tests green.

Assets 2
Loading