ABrain/Delta-NAS-Qwen2.5-Coder-7B

Delta-NAS Qwen2.5-Coder-7B-Instruct LoRA Adapter

This is a LoRA adapter for Qwen2.5-Coder-7B-Instruct, fine-tuned for delta-based Neural Architecture Search (NAS) — generating novel PyTorch image-classification architectures via unified code diffs.

Model Description

This adapter is the result of 22 iterative fine-tuning cycles on the delta-NAS pipeline described in "Delta-Based Neural Architecture Search: LLM Fine-Tuning via Code Diffs". The model generates unified diffs that modify a baseline neural network architecture to produce new, functional PyTorch models.

Training Details

• Base model: Qwen/Qwen2.5-Coder-7B-Instruct
• Fine-tuning method: LoRA (Low-Rank Adaptation)
• LoRA rank (r): 32
• LoRA alpha: 32
• LoRA dropout: 0.05
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Training cycles: 22 (iterative self-improvement)
• Total trained candidates: 793
• Admitted novel architectures: 51 (MinHash-Jaccard novelty filter + τ_acc ≥ 0.40)

Evaluation Datasets

Models were evaluated on 6 LEMUR image-classification benchmarks:

• CIFAR-10, CIFAR-100, MNIST, SVHN, ImageNette, CelebA-Gender

Key Results

Metric	Value
Trained candidates	793
Valid rate (compiles + trains)	71.8%
Mean 1-epoch accuracy	64.6% (±5.7% SD across cycles)
≥40% accuracy rate	74.5%
Best single-model accuracy	99.5% (MNIST)
Novel architectures admitted to LEMUR	51

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-7B-Instruct",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "ABrain/Delta-NAS-Qwen2.5-Coder-7B")

# Generate a diff to modify a baseline architecture
prompt = """Given the following PyTorch neural network baseline:
[baseline code here]

Generate a unified diff that creates a novel architecture variant."""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Associated Resources

• Code: ABrain-One/nn-gpt
• Generated models: ABrain-One/nn-dataset PR #204 (197 del-* prefixed architectures)
• Paper: "Delta-Based Neural Architecture Search: LLM Fine-Tuning via Code Diffs" (submitted to CVPR 2026)

Citation

@article{deltanas2026,
  title={Delta-Based Neural Architecture Search: LLM Fine-Tuning via Code Diffs},
  author={Adhikari, Santosh and Ignatov, Dmitry},
  year={2026}
}

License

Apache 2.0 License (same as the base model)