2 Commits

Author	SHA1	Message	Date
Seth	9abf9238c5	3-tier self-play: command drills, self-critique, adversarial ... Tier 1 — Command drills: Random seed prompts → generate commands → RCON validates Teaches: accurate command syntax Tier 2 — Single-shot self-critique: Model invents a tricky prompt AND responds in one call RCON validates the self-generated commands Teaches: edge-case awareness, self-evaluation Tier 3 — Adversarial self-play: Session A generates challenging prompts Fresh Session B responds cold (can't cheat) RCON validates, self-corrects on errors Teaches: robustness, generalization Usage: --tier 1|2|3|all --rounds N --focus category Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	2026-03-19 19:39:33 -04:00
Seth	c947fc3fa9	Self-play loop, Qwen3.5-9B bake-off: 70% base accuracy ... Self-play (training/scripts/self_play.py): - Model generates edge-case prompts across 9 categories - Attempts commands via RCON, self-corrects on errors - Successful traces → standard training examples - Error correction traces → multi-turn tool-calling examples - Anti-collapse: focuses on categories model is weakest in - Ready for v4 deployment, not yet active Qwen3.5-9B base model bake-off (147/1542 cases): - 70.1% OK (vs 34% Qwen3-8B base) — 2x improvement - 29.9% MISS (mostly God/prayer — no persona training) - 15.6% needed syntax fixes - Avg 7.5s response (thinking tokens) - Strong v4 candidate: better base + tool-calling architecture Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>	2026-03-19 19:35:57 -04:00