smollm3-forest-lora

A small experimental LoRA adapter that fine-tunes SmolLM3-3B to act as the genome proposer for a Pokémon Red agent trying to cross Viridian Forest to Pewter City.

Given a rollout situation (current position, sub-beats reached, catchers beaten, and the current genome), the model is trained to respond with only a JSON genome — the battle/survival parameters (hp_run_threshold, hp_heal_threshold, move-pick weights, etc.) that decide whether the agent flees or fights wild encounters and when it heals. Navigation along the route is hand-driven; the genome controls survival.

It was produced by the autotune Try → Check → Reward → Nudge loop in pcc-labs/empirical-evidence as the only step where weights are updated.

⚠️ Scope: this is a proof-of-concept adapter trained on a tiny rejection-sampled set (4 train / 1 validation example, 8 iterations). It demonstrates the pipeline; it is not a general-purpose or heavily-tuned model.

Details

Base model	EricFillion/smollm3-3b-mlx (MLX build of HuggingFaceTB/SmolLM3-3B)
Method	LoRA (fine_tune_type: lora)
Trainer	mlx_lm lora (Apple-Silicon backend)
LoRA rank / scale / dropout	16 / 16.0 / 0.05
Layers adapted	16
Learning rate	2e-5
Max sequence length	4096
Optimizer	adam
Iterations	8
Data	rejection-sampled chat SFT (system/user/assistant JSONL)

Files: adapter_config.json + adapters.safetensors (~54 MB).

Usage (MLX)

pip install mlx-lm huggingface_hub
hf download bdougie/smollm3-forest-lora --local-dir ./smollm3-forest-lora

from mlx_lm import load, generate

# Side-load the adapter against the base model
model, tokenizer = load(
    "EricFillion/smollm3-3b-mlx",
    adapter_path="./smollm3-forest-lora",
)

prompt = tokenizer.apply_chat_template(
    [
        {"role": "system", "content": "You tune a Pokemon Red agent's battle/survival genome to cross Viridian Forest to Pewter City. Respond with ONLY the JSON genome."},
        {"role": "user", "content": "This rollout reached forest beat 1 'Enter Viridian Forest' (reward 2.0 sub-beats, 1 catchers beaten, crossed=False). Current genome: {\"hp_run_threshold\": 0.2, \"hp_heal_threshold\": 0.3}"},
    ],
    add_generation_prompt=True,
)

print(generate(model, tokenizer, prompt=prompt, max_tokens=256))

To fuse the adapter into a standalone model instead of side-loading:

python -m mlx_lm fuse \
  --model EricFillion/smollm3-3b-mlx \
  --adapter-path ./smollm3-forest-lora \
  --save-path ./smollm3-forest-fused

Related

• Project: pcc-labs/empirical-evidence — the training loop that produced this adapter — and pcc-labs/pokemon-kafka, the agent it drives.
• Dataset: bdougie/pokemon-red-telemetry-sft — a broader multi-domain SFT corpus from the same project (battle-outcome, move-choice, narrator, …). Note: this adapter was trained on a separate, tiny forest-crossing genome corpus — not on that dataset, which trained a different multi-task model.

License

Inherits the base model's license (Apache-2.0).