RaifuWars-Warrior-Qwen3VL-4B

⚠️ Superseded by v2 — and the evaluation below understates one failure

Use RaifuWars-Warrior-Qwen3VL-4B-v2 instead. This model plays a recognisable game and will not win one.

tier up is the only win condition in Raifu Wars. The dataset this was trained on contained 11 examples of it, and this model declines it 137 times out of 145 offered.

The evaluation below reports tier 32/32 — 100% and did not catch that. Actions were bucketed by name prefix, which merged tier (tier-up, the winning move) with tier_stars and tier_kills (tier choice — which track to climb). Those are different actions. The bucketing averaged the refusal away, and the 85.1% headline is an average over a metric that could not see the failure.

v2 trains on 7 maps and 3,069 tier-up examples and scores 14/14 on tier-up and 98.4% on rush (against 45.5% here), with the eval reporting tier-up on its own line.

Everything else on this page is accurate as measured. It is left up as a record.

RaifuWars Warrior — Qwen3-VL-4B (merged)

Huihui-Qwen3-VL-4B-Instruct-abliterated with the RaifuWars Warrior LoRA merged in — a standalone model that plays a seat in Raifu Wars through the Warrior protocol.

85.1% agreement with the game's built-in AI on the full 841-decision held-out split (matches the model never saw), against ~20% for the base and a published prompting-only baseline of 36.3%. Legal action emitted on 841/841 decisions.

Vision is unchanged: the LoRA targeted language layers only, and the merge preserves all 315 vision tensors bit-identical to the base.

Use

model = AutoModelForImageTextToText.from_pretrained(
    "yotisstudios/RaifuWars-Warrior-Qwen3VL-4B", dtype=torch.bfloat16, device_map="auto")

Pass the Warrior payload unchanged — system prompt, board state, and the per-turn tools schema whose action_id enum is that turn's legal set. The model replies with a take_action call.

See the LoRA repo for training details, the epoch-1 alternative (better at rush, 72.7% vs 45.5%), and limitations — one map, 40 seeds, heuristic teacher, and the fact that agreement is not win rate.