Technical report: Read paper · PDF · BibTeX · All models

Collection: Spot Snake models

Spot Snake d1-2B

LiquidAI/d1-2B 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 25.32% (1,785 rows)
Best of 2 epochs, validation 99.10% (1,785 rows)
Selected checkpoint, held-out test 99.41% (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)

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
  • gguf/mmproj-f16.gguf: unmodified vision 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-d1-2B-v0.1 --local-dir models/spot-snake-d1-2B-v0.1
# then, from the tooling (see below):
python scripts/serve.py --model models/spot-snake-d1-2B-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.

One-file bundle

gguf/spot-snake-d1-2B-v0.1-bundle.gguf (1.479 GB) holds model-Q4_K_M.gguf and mmproj-f16.gguf in one file, byte for byte. The name has no quant tag or "mmproj" in it, so llama.cpp -hf, Ollama and LM Studio keep picking the separate files. Use it if you'd rather download one file. llama.cpp can't run the bundle directly, because it always needs the model and mmproj as two files. Unpack it with gguf_bundle.py (Python 3, no packages needed):

hf download Lucasgitguy-co/spot-snake-d1-2B-v0.1 gguf/spot-snake-d1-2B-v0.1-bundle.gguf gguf/gguf_bundle.py --local-dir .
python3 gguf/gguf_bundle.py unpack gguf/spot-snake-d1-2B-v0.1-bundle.gguf -o .

The script checks each file's sha256 while it unpacks. python3 gguf_bundle.py list <bundle> shows what's inside.

Checked 2026-11-16: the unpacked files have the same sha256 as the separate files in this repo.

Browser (ONNX)

onnx/decoder_q4f16.onnx (+ decoder_q4f16.onnx_data, 1.57 GB) runs in ONNX Runtime Web on WebGPU (onnxruntime-web/webgpu). It is what Spot's Biscuit Snake runs when you pick Model: d1-2B.

  • It is LiquidAI/d1-2B-ONNX decoder_q4f16 (revision 2fd6148) with only the trained part replaced: the last block's three feed-forward weights, re-quantized from this model's merged weights in the same layout (gate and up int4, down int8, blocks of 32). The graph file is LiquidAI's, unchanged. conversion/d1_patch.py makes it.
  • Inputs and outputs are LiquidAI's: one packed row (input_ids, attention_mask, position_ids, question_ids, slot_positions, empty image_features/video_features), logits at the answer slot. Rows are built by this repo's prompt.py and runner.py.
  • Checked with ONNX Runtime (CPU) through this repo's own host code (conversion/d1_eval.py): the same move as the Safetensors weights on all 32 reload-check.json rows, and 98.95% teacher agreement on the 1,708 test rows (99.41% in full precision; the gap is the 4-bit quantization of the whole decoder). See onnx-check.json.
  • The licence is unchanged: this is the same model, so LvW-license-Snake-1.0 applies.

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/d1-2B at 10e7671d44.

Raw metrics
{
  "baseline_validation": {
    "teacher_agreement": 0.25322128851540615,
    "examples": 1785
  },
  "epochs": [
    {
      "teacher_agreement": 0.9803921568627451,
      "examples": 1785
    },
    {
      "teacher_agreement": 0.9910364145658264,
      "examples": 1785
    }
  ],
  "test": {
    "teacher_agreement": 0.9941451990632318,
    "examples": 1708
  }
}
Downloads last month
236
Safetensors
Model size
2B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Lucasgitguy-co/spot-snake-d1-2B-v0.1

Finetuned
LiquidAI/d1-2B
Quantized
(16)
this model

Dataset used to train Lucasgitguy-co/spot-snake-d1-2B-v0.1

Collections including Lucasgitguy-co/spot-snake-d1-2B-v0.1