dexmac/doom-gemma3-270m-lora

doom-gemma3-270m — Gemma 3 270M LoRA for doomLaya decisions

A LoRA adapter for Gemma 3 270M, trained on the same 979 labelled questions of azalio/doomLaya v3 that were used for Laya v3 (421M encoder). It is the size-matched decoder of a comparison between "System One" decision models and fine-tuned generative LLMs; the larger sibling is dexmac/doomgemma-e2b-lora.

Code, raw results and the full test suite (Laya v3, Laya base, Laya typed-decisions, Gemma zero-shot and LoRA, rule-based oracle): https://github.com/dexmac221/doomgemma

Why it is interesting: a negative-but-useful result

130 held-out validation questions:

model	command	weapon
Gemma 3 270M, zero-shot	0.16	0.02
Laya v3	0.725	0.98
this adapter	0.89	1.00

Latency on the same RTX 4070, same 1,416 recorded game packets:

setup	p50	p99	decisions/s
Laya v3, its own server	23 ms	25 ms	41.8
this adapter, direct PyTorch (one pass, both questions batched)	22 ms	24 ms	44.7
this adapter, llama.cpp Q8_0 + grammar	71 ms	113 ms	13.2

At equal size and equal way of computing, a decoder is as fast as the Laya encoder.

But it plays worse. On FreeDoom MAP01 (seeds 48–53, 180 s) it exits only 2 of 6 games (62.8 s and 74.1 s), against 6/6 for Laya v3 and for the E2B adapter. The executor rejects its commands 60% of the time (53% "path blocked"): high per-question accuracy does not guarantee step-to-step consistency. So, with these data, there is no decoder yet that is both as fast and as good as Laya.

Limitations

• Trained and validated only on MAP01. On MAP02 nobody exits, including the rule-based teacher that produced the labels (the chain stops at a yellow-key door).
• Labels come from a rule-based teacher, not from human play.
• Base weights: trained on unsloth/gemma-3-270m-it (model.safetensors sha256 700b710a…), a mirror of google/gemma-3-270m-it. The original repository is gated, so the hashes could not be compared.

Prompt, training and credits

Same prompt format, LoRA recipe and loss (answer letter only) as dexmac/doomgemma-e2b-lora: rank 16 on q,k,v,o,gate,up,down, lr 1e-4, 5 epochs of 101 s on one RTX 4070, best epoch 5 (history in training_history.json).

Base model: Google Gemma 3 270M, distributed under the Gemma Terms of Use. Training data and harness: azalio/doomLaya, Apache-2.0 (see their NOTICE). Laya: Convai Innovations, Apache-2.0.