Lucasgitguy-co/spot-snake-d1-3B-v0.1

Spot Snake d1-3B

LiquidAI/d1-3B post-trained to pick moves in Spot's Biscuit Snake. It learns from the Spot Snake Decisions dataset: the game offers the safe moves and the model ranks them, copying a simple teacher that keeps the most open room.

Results

Teacher agreement: how often the model picks a move the teacher accepts. This is not game score, win rate or speed; those have not been measured.

Check	Teacher agreement
Base model, validation	97.82% (1,785 rows)
Best of 2 epochs, validation	99.61% (1,785 rows)
Selected checkpoint, held-out test	99.82% (1,708 rows)
Saved files reloaded, test subset	100.00% (32 rows)
model-f16.gguf, same move as Safetensors	100.00% (32 rows)
model-Q4_K_M.gguf, same move as Safetensors	100.00% (32 rows)

The first full save ran out of disk, so these weights were rebuilt by merging the saved trained adapter into the pinned base. Nothing was retrained; recovery.json records how.

Files

• model-0000N-of-0000M.safetensors: full merged weights (bf16)
• adapter/: the small rank-8 adapter that was trained (custom format, not PEFT)
• gguf/model-f16.gguf, gguf/model-Q4_K_M.gguf: llama.cpp text model
• vision projector: use mmproj-d1-3B-F16.gguf from LiquidAI/d1-3B-GGUF; training left the vision tower unchanged, and the pinned llama.cpp cannot convert this variant's projector (not tested)
• upstream runtime code and tokenizer, unchanged
• metrics.json, training.json, reload-check.json: what was measured
• LICENSE, NOTICE, licenses/: licence texts

Use

pip install "transformers>=5.22,<6" torch safetensors
hf download Lucasgitguy-co/spot-snake-d1-3B-v0.1 --local-dir models/spot-snake-d1-3B-v0.1
# then, from the tooling (see below):
python scripts/serve.py --model models/spot-snake-d1-3B-v0.1          # local game at http://127.0.0.1:8000

GGUF, with llama.cpp at revision 1b6935b or later:

llama-server -m gguf/model-Q4_K_M.gguf --port 8099
# POST /v1/systemone with a state and a choice question (see scripts/verify_gguf.py)

The scripts, dataset generator and step-by-step instructions are in the dataset repo (scripts/, INSTRUCTIONS.md). Loads with AutoModel.from_pretrained(..., trust_remote_code=True); it runs the upstream d1 code shipped in this repo.

Limits

• Ranks only the moves the game already marks as safe; keep the game's collision checks.
• The teacher is a heuristic, not an optimal player. Matching it is the whole goal here.
• Confidence scores are not calibrated after training.

Licence

LvW-license-Snake-1.0: the LFM Open License v1.0 covers the model as a whole (it is derived from Liquid AI weights, so its commercial-use limit still applies), and Apache-2.0 covers the new Spot Snake contributions. A plain MIT or Apache-2.0 licence is not possible for this model. These apply together. See LICENSE, NOTICE and licenses/. Base: LiquidAI/d1-3B at 051bcc464b.

Raw metrics

{
  "baseline_validation": {
    "teacher_agreement": 0.9781512605042016,
    "examples": 1785
  },
  "epochs": [
    {
      "teacher_agreement": 0.988795518207283,
      "examples": 1785
    },
    {
      "teacher_agreement": 0.996078431372549,
      "examples": 1785
    }
  ],
  "test": {
    "teacher_agreement": 0.9982435597189696,
    "examples": 1708
  },
  "recovery_test": {
    "teacher_agreement": 1.0,
    "examples": 32
  }
}