Modelpublic

reasoning-degeneration-dev/algo-sft-long-arithmetic-standard

sourceHugging Facemitupdated 6mo agoView on Hugging Face

0likes8downloads

Model Card

Long Arithmetic — Standard

LoRA adapter for Qwen/Qwen2.5-1.5B-Instruct fine-tuned on long arithmetic via Algorithmic Template SFT.

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	Algorithmic Template SFT
Framework	LLaMA-Factory (SFT stage)
LoRA rank	64
LoRA target	all linear layers
Learning rate	1e-4
Epochs	3
Batch size	4 (grad accum 4)
Cutoff length	32,768 tokens
Training data	5,000 deterministic carry-propagation traces (d4: 3-digit x 2-3 digit multiply)

Evaluation (v3, MAX_TOKENS=32768)

Split	Accuracy
Test (in-distribution)	92.6%
Harder variant	21.2%
Structural OOD	0.0% (chain operations)

Notes

Strong in-distribution but collapses on structural OOD (chain operations). Neither algo nor distill generalizes here.

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-standard")
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)