---

license: other

license_name: lfm1.0

license_link: https://huggingface.co/LiquidAI/d1-3B/blob/da1fe36a861f24690f27f622dca1d8688503d113/LICENSE

base_model: heath0xFF/touch-grass-v0

base_model_relation: quantized

language:

• en

tags:

• gguf
• llama-cpp
• d1
• decision-model
• experimental

---

touch-grass v0 — GGUF Q8_0

Antidoom, but for agents. A text-only Q8_0 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.

This is a decision model, not a chat model that generates action labels.

Use llama.cpp's native /v1/systemone API. Ordinary chat clients are not

necessarily compatible.

Run

Use a d1-capable llama.cpp build. Tested with 0.7.0-dev, commit

10b8b7553494f3638da7668cf17cac089229122a,

on Apple M5 Pro with Metal. Older builds without lfm2-d1 support will fail.

With the Hugging Face CLI installed:

hf download heath0xFF/touch-grass-v0-GGUF --revision v0 --local-dir touch-grass-v0-GGUF
llama-server -m touch-grass-v0-GGUF/touch-grass-v0-Q8_0.gguf \
  --host 127.0.0.1 --port 18081 --ctx-size 2048 --parallel 1 \
  --batch-size 2048 --ubatch-size 2048 --no-context-shift --no-webui -ngl 99

In another terminal:

curl --fail-with-body http://127.0.0.1:18081/v1/systemone \
  -H 'Content-Type: application/json' \
  --data-binary @touch-grass-v0-GGUF/request.json

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 response includes choice, all five

native candidate probabilities, and zero output tokens. Do not interpret native

confidence fields or probabilities as calibrated real-world safety estimates.

Only text was exported: no vision tower or multimodal projector is included.

Keep the complete rendered input within v0's 2,048-token limit. The companion

project's PyTorch CLI does not load GGUF; use the native endpoint above.

Do not expose this unauthenticated example server beyond loopback.

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 0.0413 on test and **0.0757

on gold**. Mean absolute changes are 0.000266 and 0.001392 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.

Raw predictions and metrics

· Conversion and evaluation recipe

· Original training and limitations

Provenance and license

Source release: heath0xFF/touch-grass-v0@8cd6d51892d82b61db40f2ff4889a7dfe0755f89.

Converted through F16 GGUF, then quantized with llama-quantize Q8_0. No

retraining, calibration dataset, replacement head, or generated-label scoring.

The GGUF includes the native d1 decision type and System One prompt template.

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.