Qwen3-4B-Instruct · MegaGem distilled (paraphrase) — weights-only champion
Merged weights: djdumpling/qwen3-4b-instruct-megagem-sft-step1200-v2 (SFT blueprint) + the round-1 paraphrase distillation LoRA (distill1a_d1a_para). SFT + distillation are baked in; serve with plain vLLM/transformers.
Trained via ExIt with an analytic expert: self-play games → a fitted opponent bid-distribution price law (F̂) + a self-play value head (V̂) → closed-form expected-surplus best-response labels on deviation turns → self-paraphrased in the blueprint's own voice → SFT LoRA. No RL gradient; no opponent-API data in any learned component (the price law is fit on the blueprint's OWN self-play bids).
Held-out vs 2×Gemini-3-Flash (paired CRN, weights-only, NO test-time selector): 29.3% → 40.7% win-rate, CV-adjusted +7.4 score/game (t≈3.0). The deployable champion adds a zero-LLM-cost analytic EV selector at test time (≈ +11–14 vs Flash/Pro/Sonnet) — that selector is code, not these weights.