suryadv/IncorrectTraceSFT-Qwen2.5-7B-MATH-random-mask-80

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The suryadv/IncorrectTraceSFT-Qwen2.5-7B-MATH-random-mask-80 is a 7.6 billion parameter language model, fine-tuned from Qwen2.5-7B, specifically for mathematical reasoning tasks. Developed by suryadv, this model explores the impact of using incorrect traces in supervised fine-tuning, focusing on a 'random_mask_80' condition within a math masking experiment. It is designed for research into mathematical reasoning and the effects of SFT data selection strategies.

Loading preview...

Model Overview

The suryadv/IncorrectTraceSFT-Qwen2.5-7B-MATH-random-mask-80 is a 7.6 billion parameter model, fine-tuned from the Qwen2.5-7B base model. It is a product of research into how incorrect traces should be utilized in Supervised Fine-Tuning (SFT), specifically within the context of mathematical problem-solving.

Key Characteristics

  • Mathematical Reasoning Focus: This model is a specific checkpoint from an experiment investigating SFT techniques for improving mathematical reasoning capabilities.
  • Incorrect Trace SFT: It represents the "random_mask_80" condition from a "math masking" experiment, exploring the impact of masking parts of incorrect reasoning traces during training.
  • Training Details: The model was trained using SLIME with Megatron-LM, BF16, a global batch size of 64, and AdamW optimizer with a peak learning rate of 5e-6, over 125 optimizer updates.
  • Research-Oriented: Released primarily for research purposes, particularly for studying mathematical reasoning and the effectiveness of different SFT data selection strategies.

Intended Use

This model is suitable for:

  • Research on SFT Data Selection: Investigating the effects of including or modifying incorrect reasoning traces in supervised fine-tuning datasets.
  • Mathematical Reasoning Studies: Exploring how models learn and perform on mathematical tasks under specific training conditions.
  • Comparative Analysis: Evaluating the performance of this specific SFT condition against other experimental setups detailed in the associated code repository.