tianzl66/Qwen2.5-7B-Instruct-InstructionFollowing-SpectralSurgery-HNS8p2

Qwen2.5-7B-Instruct + Instruction Following + Spectral Surgery

This repository contains a Spectral Surgery adapter derived from the instruction-following LoRA checkpoint of Qwen2.5-7B-Instruct.

Post-hoc Spectral Surgery applies Hybrid Newton-Schulz (HNS) editing directly to the LoRA adapter. No additional gradient-based training is performed during Spectral Surgery.

Base Model

Qwen/Qwen2.5-7B-Instruct

Source LoRA

• Dataset: Tulu-3 SFT instruction-following split
• Samples: 29,980
• Epochs: 2
• Sequence length: 4096
• Global batch size: 128
• LoRA rank: 16
• LoRA alpha: 32
• Learning rate: 2e-4
• Seed: 42

Spectral Surgery

• Target: all LoRA modules
• Fast HNS steps: 8
• Stable HNS steps: 2
• Output rank: 16
• Nuclear norm: preserved
• See spectral_edit_meta.json for exact edit metadata.

Evaluation

Evaluation on IFEval (541 prompts, 834 instructions).

Model	Prompt Strict	Prompt Loose	Instruction Strict	Instruction Loose
Base	65.43%	70.43%	75.18%	79.38%
LoRA SFT	64.70%	69.32%	73.38%	77.82%
Spectral Surgery HNS 4+1	69.69%	75.05%	77.94%	82.49%
Spectral Surgery HNS 8+2	69.13%	73.94%	77.70%	81.65%

Relative to the vanilla LoRA checkpoint, Spectral Surgery HNS 8+2 improves:

• Prompt Strict by 4.44 percentage points
• Prompt Loose by 4.62 percentage points
• Instruction Strict by 4.32 percentage points
• Instruction Loose by 3.84 percentage points

The edited adapter also exceeds the base model on all four IFEval metrics.

Settings: tokenizer chat template, greedy decoding, max_new_tokens=2048, vLLM backend, max model length 4096, seed 42.

Files

• adapter_model.safetensors: PEFT LoRA weights
• adapter_config.json: PEFT configuration
• eval-ifeval/metrics.json: aggregate IFEval metrics
• eval-ifeval/outputs.jsonl: per-prompt generations and instruction results
• spectral_edit_meta.json: exact HNS edit metadata