touch-grass v0 — MLX 8-bit

Antidoom, but for agents. An Apple Silicon export of the merged touch-grass v0 checkpoint, fine-tuned from Liquid AI d1-3B. It scores five next actions:

CONTINUE · RETRY · CHANGE_STRATEGY · ESCALATE · TOUCH_GRASS

Experimental and advisory only. Not a safety gate or authorization system. Use mlx-vlm's native d1 decision methods, not text generation or a generic chat interface. No replacement classifier head is added.

Run on Apple Silicon

Tested on Apple M5 Pro with MLX 0.32.4 and mlx-vlm 0.7.8, commit 73d366ecf9233b3fc6549531a7f2386726d09ff6. mlx-vlm versions without d1 support will not work.

Save the following as judge_mlx.py:

import json
from pathlib import Path

from huggingface_hub import snapshot_download
from mlx_vlm import load

path = Path(snapshot_download("heath0xFF/touch-grass-v0-MLX-8bit", revision="v0"))
model, processor = load(path)
request = json.loads((path / "request.json").read_text())
question = request["questions"]["action"]
rows, tokens = model.probabilities(processor, request["state"], [question])
probabilities = dict(zip(question["criteria"], rows[0], strict=True))
print({"action": max(probabilities, key=probabilities.get), "probabilities": probabilities})

Run with uv:

uv run --with 'mlx-vlm @ git+https://github.com/Blaizzy/mlx-vlm.git@73d366ecf9233b3fc6549531a7f2386726d09ff6' \
  --with 'mlx==0.32.4' --with 'transformers==5.19.0' python judge_mlx.py

request.json includes a demo task, chronological trajectory, and the exact five-action question. Replace its state to judge another trajectory; retain the question for comparable behavior. The code uses the native, unrounded model.probabilities readout. The higher-level mlx_vlm.predict API also works, but rounds its probabilities to four decimal places and calls the selected label value. Neither API generates output tokens for these decisions.

Keep the complete rendered input within v0's 2,048-token limit; the native MLX method does not enforce this project-specific limit. The evaluation script checks it. The companion project's PyTorch CLI does not load these MLX weights.

Measured checks

Compared with the published CUDA BF16 checkpoint, using the same input files and per-case candidate orders. Both are held-out sets, not quantization data.

Set	Accuracy	Macro F1	Stop recall	False stops	Same top action as BF16
Synthetic test: 150 rows / 10 families	90.0%	0.893	30/30	0/90	150/150
Gold: 30 separately authored cases	90.0%	0.898	6/6	1/18	30/30

The largest candidate-probability change is 0.0371 on test and 0.1045 on gold. Mean absolute changes are 0.000351 and 0.002098 respectively. Differences include runtime, precision, batch shape, and quantization effects; this is not a measurement of quantization alone. Matching top actions does not mean matching probabilities or calibration. The original three gold errors remain.

False stops count TOUCH_GRASS predictions on true CONTINUE, RETRY, or CHANGE_STRATEGY cases. The gold set is assistant-authored, not human-adjudicated. Synthetic variants are correlated. No real-agent safety validation was done. Never treat a high probability as permission to repeat a payment or take a destructive action. No post-training probability calibration was performed.

Raw predictions and metrics · Conversion and evaluation recipe · Original training and limitations

Provenance and license

Source release: heath0xFF/touch-grass-v0@8cd6d51892d82b61db40f2ff4889a7dfe0755f89. Converted with mlx-vlm: 8-bit affine quantization, group size 64, with other tensors in float16. Language linear/embedding weights are quantized; the vision tower and projector remain float16. Vision weights are retained but vision use was not tested. No retraining or calibration dataset was used.

The upstream PyTorch modules are not bundled. config.json keeps its D1Model marker so mlx-vlm selects its built-in native d1 implementation. These are MLX weights, not a checkpoint to load through Transformers. export.json records the tool revisions, source hash, and artifact hashes.

Weights use the LFM Open License v1.0, including its commercial-use restrictions. LICENSE and NOTICE are included. The companion code's MIT license does not relicense the weights. Not affiliated with or endorsed by Liquid AI.