Modelpublic

reasoning-degeneration-dev/algo-sft-long-arithmetic-distill-qwq

sourceHugging Facemitupdated 6mo agoView on Hugging Face

0likes6downloads

Model Card

Long Arithmetic — QwQ Distillation

LoRA adapter for Qwen/Qwen2.5-1.5B-Instruct fine-tuned on long arithmetic via QwQ-32B Distillation.

Part of the Algorithmic SFT vs Distillation experiment studying whether deterministic algorithmic templates teach procedural reasoning more effectively than distillation from large reasoning models.

Training

Parameter	Value
Base model	Qwen/Qwen2.5-1.5B-Instruct
Method	QwQ-32B Distillation
Framework	LLaMA-Factory (SFT stage)
LoRA rank	64
LoRA target	all linear layers
Learning rate	1e-4
Epochs	3
Batch size	1 (grad accum 16)
Cutoff length	32,768 tokens
Training data	5,000 QwQ-32B reasoning traces (d4, filtered). Teacher solve rate: 43.8%

Evaluation (v3, MAX_TOKENS=32768)

Split	Accuracy
Test (in-distribution)	90.6%
Harder variant	8.4%
Structural OOD	6.8%

Notes

Nearly tied with algo SFT in-distribution (90.6% vs 92.6%). Slight OOD edge (6.8% vs 0%) but both effectively fail.

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
model = PeftModel.from_pretrained(base, "reasoning-degeneration-dev/algo-sft-long-arithmetic-distill-qwq")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")

Related Datasets

—Training data (63K algo traces)
—Distillation data (24K QwQ traces)
—Eval results (aggregate scores)
—Eval questions (11K test/val/harder/OOD)