RaifuWars-Warrior-Qwen3VL-4B-LoRA

⚠️ Superseded by v2 — and the evaluation below understates one failure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 LoRA — Qwen3-VL-4B

A LoRA that lets a 4B model play a seat in Raifu Wars through the Warrior protocol, trained on RaifuWars-Warrior-Experimental-SFT — 9,030 decisions by the game's built-in heuristic AI, each paired with the board state and the complete legal action list.

The dataset exists because prompting did not work: told in plain English what the built-in AI would do on that exact board, a 4B model complied 13% of the time on fight and took rush 0 times out of 1,258 offered. Hinting raised overall agreement from 36.3% to 40.1% and converted to zero wins in 40 matches. "A hint is a suggestion and can be declined. A trained target cannot."

Measured

200 held-out decisions, greedy decoding. The validation split is divided by match, so these are boards from games the model never saw.

Against the dataset card's published prompting-only baseline of 36.3% (40.1% with hints).

Per action type, with counts, because the training mix is 36% move and 23% end turn and a model that collapsed onto those two would still score respectably overall:

The rush result is the point. The dataset card reports a prompted model taking rush 0 of 1,258 times offered; the untuned base here reproduces that exactly at 0/21. Both tuned checkpoints break it.

Epoch 1 is included in epoch1-alternate/ because the two checkpoints trade differently. Epoch 2 wins overall and fixes an attack regression that epoch 1 introduced (36.4% → 95.5%), but epoch 1 is markedly better at rush (81.0% vs 52.4%). If rush is the behaviour you care about, use epoch 1.

Sample sizes are small per category. rush is 21 decisions, reload 7, tier 8. The epoch-1-vs-2 rush gap is 17/21 against 11/21 — suggestive, not settled. A full 841-row run was attempted and is not reflected here; treat the per-action numbers as directional.

Legality was never the problem. The base model already emitted a legal action_id on 200/200 decisions. This adapter does not teach valid syntax — it changes which legal action gets chosen.

Usage

Serve with llama.cpp, adapter GGUF included:

llama-server -m Qwen3VL-4B-abliterated-f16.gguf \
             --lora-scaled raifu-warrior-lora-f16.gguf:1.0 \
             -ngl 99 -c 32768 --jinja -fa on --port 8081 --host 0.0.0.0

Then POST the Warrior payload unchanged — system prompt, board, and the per-row tools schema whose action_id enum is that turn's legal set:

requests.post("http://localhost:8081/v1/chat/completions", json={
    "messages": msgs, "tools": row["tools"], "temperature": 0, "max_tokens": 128,
})
# -> {"tool_calls": [{"function": {"name": "take_action",
#                                  "arguments": "{\"action_id\": \"move_2_9\"}"}}]}

Or with peft:

model = AutoModelForImageTextToText.from_pretrained(
    "huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated", dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, "yotisstudios/RaifuWars-Warrior-Qwen3VL-4B-LoRA")

Note AutoModelForImageTextToText, not AutoModelForCausalLM — this is a Qwen3VLForConditionalGeneration.

Vision is untouched

The base is a vision-language model and the training data is text-only. The seven targeted projections match 252 language modules and 0 vision modules — the ViT uses qkv, linear_fc1, linear_fc2. The resulting adapter has 504 tensors, none of them vision, all under model.language_model.layers.*. Image understanding is bit-identical to the base, by construction rather than by care.

Training

Loss 0.742 → 0.017; held-out eval loss 0.021, below train throughout. Low absolute loss is expected and not very informative here — the target is one short action_id from a constrained enum, so most tokens are trivially predictable. Agreement is the metric that distinguishes learning from format-fitting.

Two build details that matter:

The target text is derived, not written. The conversation is rendered twice through the chat template — once without the assistant turn, once with — and the difference is the training target. Whatever the template emits for a tool call is what the model learns to emit. Hand-writing <tool_call> blocks would risk a train/serve mismatch that nothing reports. (The dataset's content: null on tool-calling assistant turns must be replaced with "" first, or Qwen3-VL's jinja raises TypeError: 'NoneType' object is not iterable.)

Over-length rows are dropped, never truncated. The trainer truncates long prompts from the right, which here would cut off the legal-action list and the closing instruction — training the model to name an action_id it was never shown. 20 of 9,050 rows exceeded 8192 tokens and were dropped. A 4096 cap would have dropped 2,607; a 2048 cap would have dropped every row.

build_dataset.py, train.py and eval_actions.py are included.

Limitations

Inherited from the data, and they are real:

• One map. Every training row is Arboretum. Tile preferences learned here may not transfer, and validation cannot detect that because it shares the map.
• 40 seeds is the true sample size, not 9,030. Decisions within a match are strongly correlated.
• The teacher is a heuristic, not optimal. This is imitation of a decent player. 84% agreement is 84% agreement with that, and the ceiling is its skill.
• Agreement is not win rate. Nothing here measures whether the model wins games. The prompting baseline reached 40.1% agreement and won zero of 40 matches, which is exactly why agreement alone should not be trusted as a proxy. Run it in a real match.
• last_action is absent. At inference the sidecar appends a line describing how the previous action resolved mid-turn; the traces do not contain it, so the model never saw it in training. A known train/serve mismatch, inherited and left explicit.

Licence

Apache-2.0, matching the base model. Note the training data is GPL-3.0 — whether a dataset licence reaches model weights is unsettled, and this repo takes the position that it does not. Users with a different view should treat the weights as GPL-3.0 derived.