dexmac/doomgemma-e2b-lora

doomGemma — Gemma 4 E2B LoRA for doomLaya decisions

A LoRA adapter that turns Gemma 4 E2B into a typed decision model for the doomLaya FreeDoom agent. It was trained on exactly the same 979 labelled questions that were used to train Laya v3 (a 421M "System One" encoder), so the two can be compared on equal data.

The model does not play by itself. The doomLaya agent turns the game state into text (health, weapons, visible enemies with ids and distances, items, doors) and asks two multiple-choice questions, command and weapon; an executor aims, plans paths and presses buttons. All choices belong to the model.

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

Results (RTX 4070, 25 Sep 2026)

130 held-out validation questions of doomLaya v3:

model	command	weapon	ECE (raw)
Gemma 4 E2B, zero-shot	0.51	0.78	0.38
Laya v3 (azalio/laya-doom-v3)	0.725	0.98	0.06
this adapter (Q8_0 GGUF, via llama.cpp)	0.96	1.00	0.04

FreeDoom MAP01, skill 3, 180 s limit, seeds 48–53 (none used for training or validation):

player	exits	mean time	range	deaths
Laya v3	6/6	59.5 s	42.9–71.0 s	0
this adapter	6/6	60.0 s	56.6–64.2 s	0
rule-based teacher that produced the labels	6/6	81.6 s	64.5–124.9 s	1

The models exit faster than the rule-based teacher because they imitate it imperfectly and pick up fewer items; the teacher does not optimise time-to-exit.

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

setup	p50	p99
Laya v3, its own server	23 ms	25 ms
this adapter, llama.cpp Q8_0 + grammar	59 ms	118 ms
this adapter, direct PyTorch bf16	84 ms	100 ms

Limitations

• Trained and validated only on MAP01. On MAP02 nobody exits (0/3), including the rule-based teacher: the chain stops at a yellow-key door (sector 37, linedef special 27), which doomLaya offers as an openable door while the key is far away and out of sight. Generalisation to new maps cannot be measured on this bench.
• The labels come from a rule-based teacher, not from human play: the adapter learns to imitate those rules.
• The validation set is small (130 questions); calibration numbers carry noticeable noise.

How it is used

Each option becomes a letter. A grammar allows a single token among the valid letters, and the option probabilities are read from the logprobs of that token (one prefill + one token, no free-form generation). The exact prompt:

LETTERS = "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
SYSTEM = ("You control a DOOM player in real time. "
          "Pick the best option for each question. Reply only with the letters.")

def build(state, name, question):            # one question per request
    keys = list(question["criteria"])
    lines = [f"{LETTERS[i]}) {question['criteria'][k]}" for i, k in enumerate(keys)]
    user = (state + "\n\n" + f"{name.upper()} — {question.get('instructions', '')}\n"
            + "\n".join(lines) + f"\n\nAnswer with 1 letter, in this order: {name.upper()}.")
    return [{"role": "system", "content": SYSTEM}, {"role": "user", "content": user}], keys

The chat template is applied with add_generation_prompt=True, enable_thinking=False; the answer is the next token.

Training

• Data: training/v3/train.json (979) and validation.json (130) from azalio/doomLaya at commit b25edd3 (Apache-2.0). Game seeds 42/43 for training, 44 for validation, MAP01 only.
• LoRA rank 16, alpha 32, dropout 0.05 on q,k,v,o,gate,up,down of the language model; AdamW, lr 1e-4, weight decay 0.01, gradient accumulation 8, clip 1.0, seed 771. Loss only on the answer letter. Best epoch by validation accuracy (command + weapon): epoch 2 of 3, 550 s per epoch on one RTX 4070 (history in training_history.json).
• Base weights: google/gemma-4-E2B-it (model.safetensors sha256 2db5482b…).

Credits and licences

• Base model: Google Gemma 4 E2B. This adapter is a Gemma derivative and is distributed under the Gemma Terms of Use.
• Training data and game harness: azalio/doomLaya, Apache-2.0 (see their NOTICE).
• Laya: Convai Innovations, Apache-2.0.
• FreeDoom and ViZDoom under their respective licences.