jevons-lfm25-1.2b-systemone

jevons-lfm25-1.2b-systemone

Seed LoRA for jevons: a local System One server that scores allowed outcomes from logits and assembles {choice, probabilities, confidence, noul, score, legend} in code.

This is not official Jev. Official Jev cannot be cloned. These weights only bias LFM2.5 toward the same option-key / yes-no / score-digit tokens jevons reads at serve time.

Base model

QLoRA on the MLX 8-bit checkpoint (rank 16, scale 2.0, 16 layers, attn q/k/v/o + MLP w1/w2/w3). Prompt tokens were masked. Completions are teacher-forced aliases that match serve-time scoring: option keys, yes/no, level digits.

Serve with jevons

Install jevons from github.com/gopalanj/jevons, then download this adapter and the 8-bit base:

cd jevons
uv sync --extra mlx --extra dev

uv run hf download mlx-community/LFM2.5-1.2B-Instruct-8bit \
    --local-dir models/LFM2.5-1.2B-Instruct-8bit

uv run hf download gopalanj/jevons-lfm25-1.2b-systemone \
    --local-dir adapters/lfm25-1.2b-systemone

Serve at T=1 with calibration off (do not apply a leftover calibration.json from the base model):

JEVONS_ADAPTER=adapters/lfm25-1.2b-systemone \
JEVONS_TEMPERATURE=1 \
JEVONS_CALIBRATION=off \
JEVONS_MODEL=models/LFM2.5-1.2B-Instruct-8bit \
uv run jevons serve --host 127.0.0.1 --port 8000

Or:

uv run jevons serve \
  --adapter adapters/lfm25-1.2b-systemone \
  --temperature 1 \
  --calibration off

POST /v1/systemone matches the TypeSafe HTTP contract. The adapter autoloads from adapters/lfm25-1.2b-systemone when adapters.safetensors is present. Disable with --adapter off.

Honest metrics (T=1, no calibration)

Seed-only official-Jev teacher aliases (78 train items / 126 examples; frozen holdout 17 items / 29 examples). Schema validity is 100% because jevons never asks the model to write JSON.

Full-set modal 83.9% includes the 78 train items. The honest ship metric is holdout modal 72.4%, unchanged vs base and below a 90% bar. Holdout ECE got worse (0.122 → 0.248). This is a ranking adapter, not calibrated System One / RLCD.

Training

• Data: seed teacher aliases only (not grown templates)
• Iters: 64 (2 epochs), batch 4, AdamW 5e-5, ~131s on Apple Silicon
• Failed grown run is not these weights: repetitive grow.py templates + LoRA scale 20 collapsed choice (holdout modal 38%). Do not serve that run.

See hyperparams.json / hyperparams.md in this repo.

License caveat

• Adapter + LFM weights: LFM 1.0 (license: other / lfm1.0). Obtain and accept a license from Liquid AI before downloading or using the base model or this derivative adapter. This repo does not ship LFM base weights.
• jevons server code: MIT — github.com/gopalanj/jevons

Files

• adapters.safetensors — MLX LoRA weights
• adapter_config.json — mlx-lm LoRA config
• hyperparams.json / hyperparams.md — training report and eval numbers