Qwen-2.5-3B-instruct-bioaligned-qlora

QLoRA adapter weights for a bioaligned fine-tune of Qwen/Qwen2.5-3B-Instruct.

Note: This repository contains only the LoRA adapter weights, not the full model. You must have access to the base model to use this adapter.

Merged model: Bioaligned/Qwen-2.5-3B-Instruct-Bioaligned

Organization: Bioaligned Labs (nonprofit)

Paper: (https://arxiv.org/abs/2603.09154)

Model Description

This adapter shifts model preference toward biological information sources when evaluating engineering problems--a property we call bioalignment. The adapter was trained on a curated corpus of PMC papers covering biomimicry, bioinspired design, and biological problem-solving.

Quick Start

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-3B-Instruct",
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load adapter
model = PeftModel.from_pretrained(
    base_model,
    "Bioaligned/Qwen-2.5-3B-instruct-bioaligned-qlora"
)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-3B-Instruct")

# Generate
inputs = tokenizer("Your prompt here", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Parameter	Value
Base model	Qwen/Qwen2.5-3B-Instruct
Method	QLoRA (4-bit NF4 quantization)
LoRA rank	16
LoRA alpha	32
Target modules	All attention and MLP layers
Learning rate	1e-5
Epochs	3
Training format	Instruction-tuned only
Corpus	~6M tokens from PMC Open Access papers

Note: Trained on instruction-formatted data only (no continued pretraining mix), as the mixed format used for Llama was incompatible with Qwen.

Evaluation Results

Bioalignment Benchmark (50 prompts across materials, energy, manufacturing, algorithms):

Metric	Base	Bioaligned	Change
Delta p_up (valence)	-0.111	-0.056	+51%

No capability degradation on standard benchmarks (MMLU, HellaSwag, ARC, WinoGrande).

Limitations

• Adapter only; requires base model access
• 51% improvement (vs. 93% for Llama) due to instruction-only training
• Trained on 3B model; scaling behavior unknown
• Measures stated probabilities, not downstream behavior

Citation

[TODO: Add citation when paper is published]

License

This adapter is released under the Apache 2.0 License, consistent with the base Qwen 2.5 model license.

---

Bioaligned Labs -- AI safety research