A newer, better version is out: BioDecision v2 LoRA adapter. BioDecision v2 is trained on 1.75M decisions from 57 sources (22 new) and scores 75.3% vs 69.4% for BioDecision v1 on 63,432 held-out biomedical test decisions. This BioDecision v1 page stays available.
BioDecision-4B LoRA adapter (v1.1)
LoRA weights of BioDecision-4B, for use on top of Qwen/Qwen3.5-4B (revision 851bf6e). The merged model is the simpler choice; results, usage and training details are on its card.
| Rank / alpha | 16 / 32, all linear layers incl. Gated-DeltaNet projections |
| Training | stage 1: 1.08M decisions, 16,880 steps; stage 2: 32k-decision patch, 505 steps |
| Calibration | divide option-letter logits by T = 1.3929 |
import torch
from transformers import AutoModelForCausalLM
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-4B", revision="851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a", dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base, "lighteternal/biodecision-tev1-4b-lora-v1.1").merge_and_unload()
Stage-1 weights (v1.0) are available merged: revision="v1.0" on lighteternal/biodecision-tev1-4b.
- Downloads last month
- 46
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support