xiaoyuchen1/Qwen3-1.7B-OPD-fwdKL-from-Qwen3-8B-GRPO

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 27, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The xiaoyuchen1/Qwen3-1.7B-OPD-fwdKL-from-Qwen3-8B-GRPO model is a 1.7 billion parameter language model distilled from xiaoyuchen1/Qwen3-8B-GRPO-MATH using on-policy distillation (OPD) with a forward-KL objective. This model is specifically optimized for solving MATH-style problems, serving as a control arm in a study on OPD signal improvement. It excels at mathematical reasoning tasks, particularly those found in the MATH training distribution.

Loading preview...

Model Overview

This model, xiaoyuchen1/Qwen3-1.7B-OPD-fwdKL-from-Qwen3-8B-GRPO, is a 1.7 billion parameter student model distilled from the larger xiaoyuchen1/Qwen3-8B-GRPO-MATH teacher model. The distillation process utilizes On-Policy Distillation (OPD) with a forward-KL objective over the teacher's top-128 support. This approach ensures the student model learns from states it actually visits, pushing it to cover all of the teacher's modes rather than collapsing onto one.

Key Characteristics

  • Distillation Method: On-policy distillation (OPD) with a forward-KL objective.
  • Teacher Model: xiaoyuchen1/Qwen3-8B-GRPO-MATH (8 billion parameters).
  • Student Initialization: Qwen/Qwen3-1.7B.
  • Training Data: Fine-tuned on the xiaoyuchen1/opd-math-data dataset, specifically 7,496 MATH training problems.
  • Objective: Aims to reduce the KL divergence between the student and teacher on the student's own trajectories, focusing on mass-covering to retain teacher's modes.

Performance and Use Cases

This model is primarily designed and evaluated for MATH-style problem-solving. Benchmarks show strong performance on math500 (e.g., 74.42 avg@16), which aligns with its training distribution. While also evaluated on aime24 and aime25, high truncation rates due to the 4,096-token output limit mean math500 is the most reliable indicator of its mathematical reasoning capabilities. It is important to note that this model is a control arm in a research study and has not undergone safety or alignment tuning.

Limitations

  • Trained with only 150 optimizer updates, a shorter run than some comparable studies.
  • Evaluated exclusively on MATH-style problems in a non-thinking chat format.
  • No safety or alignment tuning has been performed.

Finetunes
1 models