lighteternal/biodecision-v2-4b-lora

BioDecision-4B v2 LoRA adapter

LoRA weights of BioDecision-4B v2, 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 including Gated-DeltaNet projections
Training	one epoch from the base model: 27,347 steps × 64 = 1,750,208 decisions of the 1,750,239-decision pool (final incomplete batch dropped)
Calibration	divide option-letter logits by T = 1.0614, or by the per-family value in temperatures.json

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-v2-4b-lora").merge_and_unload()

BioDecision family

	BioDecision v2 (this release)	BioDecision v1 (previous)
Full model: merged weights, load and run	biodecision-v2-4b	biodecision-tev1-4b
LoRA adapter for Qwen/Qwen3.5-4B	biodecision-v2-4b-lora	BioDecision v1 LoRA adapter
Training data, test sets and overlap masks	biodecision-v2-data	BioDecision v1 training data + part 2
Live demo	biodecision-demo