Modelpublic

lafalce/system-one-model

sourceHugging Faceapache-2.0updated 20d agoView on Hugging Face

0likes

Model Card

System One (phase2 student)

Local open-weight decision model: JSON or text state in, typed choice / score / noul out. One forward pass, no text generation.

This Hub repo is the ship checkpoint from mateolafalce/system-one-model: ModernBERT-base + LoRA r=16 + a 2-layer fp32 scoring head. Load it with that repo's StudentModel.from_pretrained.

Files

Path	What
`backbone/`	PEFT LoRA adapter on `answerdotai/ModernBERT-base`
`head.pt`	Decision head (fp32)
`student.json`	Backbone name, head depth, val metrics
`tokenizer/`	Tokenizer snapshot
`temps.json`	Temperature scaling (`noul` K=2 T=0.6; `score` K=3–5 T=0.7)

Test numbers (proof gold)

Task	Acc	ECE
BANKING77 (77-way)	88.5%	0.059
BANKING77 coarse (8-way)	95.4%	0.016
SMS spam	98.9%	0.009
SST-5	55.9%	0.033

BANKING77 missed the 90% ship gate by 1.5 pt. The frozen teacher (Qwen2.5-7B-Instruct-AWQ, letter-logit scoring) ceilings at 56% on that task. Do not scale the student to chase 90%.

Serve

From the GitHub repo, with this folder as --ckpt:

bash

python scripts/07_serve.py --ckpt . --temperatures temps.json --port 8010

POST /v1/systemone with a state and typed questions. Cap is 512 tokens. English only.

Hardware

Fits an 8 GB RTX 3070 (~0.6 GB VRAM at serve). Do not load the Qwen teacher on the same GPU at the same time.