nicolasembleton/openjev-minicpm5-2b-lora-phase1

Open-Jev Phase-1 LoRA — MiniCPM5-2B

PEFT LoRA adapter for openbmb/MiniCPM5-2B, fine-tuned for closed-set first-token scoring on Open-Jev Choice / Noul / Score fields (Bev/Jev-style: one forward pass, candidate-token logits; not free-form generation).

Results (frozen panel, never in train)

Slice	Metric	Zero-shot	This LoRA	Δ
Test n=900	overall accuracy	0.467	0.702	+23.6 pp
Test	Noul recall	0.118	0.735	+61.8 pp
OOD n=300	overall accuracy	0.473	0.710	+23.7 pp

Scorekeeper planted gold; see local results/openjev-phase1/RUN_CARD.md for full tables.

Recipe

• Base: openbmb/MiniCPM5-2B
• Train: stratified Open-Jev subset, n=6000 (2000 Choice / 2000 Noul / 2000 Score); dataset rev 10ad6888333fa97f8c948192797bad3de3040802
• Objective: causal LM loss only on the gold index-surrogate token after Answer:
• LoRA: r=16, α=32, dropout=0.05; targets q/k/v/o/gate/up/down_proj
• Train: 1 epoch, lr=2e-4, effective batch 16, max length 2048, bfloat16
• Hardware: Modal A10G (~59.5 min wall, ~15.2 GB peak VRAM)
• Seed: 20260924

Load

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "openbmb/MiniCPM5-2B"
adapter_id = "nicolasembleton/openjev-minicpm5-2b-lora-phase1"

tok = AutoTokenizer.from_pretrained(base_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(base_id, trust_remote_code=True, torch_dtype="auto")
model = PeftModel.from_pretrained(model, adapter_id)

For Open-Jev TypeSafe-compatible scoring, point OPENJEV_ADAPTER_PATH at this adapter (or a local checkout) on top of the same base.

Browser / WebGPU note

This repo is the PEFT adapter only. In-browser Transformers.js needs a merged weights export (and ideally quantized ONNX). That path is separate from this Hub upload.

Intended use

Research and demos of closed-set Choice / Noul / Score scoring on Open-Jev-style prompts. Not a general chat model. Do not treat closed-menu argmax as ground truth without an external scorekeeper.

License

Adapter weights follow the base model license terms for derivatives of MiniCPM5-2B (Apache-2.0 style redistribution where permitted by the base). Training data: Open-Jev v1.1.