andresnowak/Qwen3-0.6B-MNLP_mcqa_rl

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:May 31, 2025Architecture:Transformer Featherless Exclusive Warm

The andresnowak/Qwen3-0.6B-MNLP_mcqa_rl model is a 0.8 billion parameter Qwen3-based language model developed by andresnowak, fine-tuned for multiple-choice question answering (MCQA). It was trained using Reinforcement Learning (RL) with the GRPO method, specifically rewarding outputs that follow a '[Letter]. [Answer]' format. This model demonstrates strong performance on various STEM-related MCQA benchmarks, particularly excelling in Math QA tasks, and supports a 32768 token context length.

Loading preview...

Overview

This model, andresnowak/Qwen3-0.6B-MNLP_mcqa_rl, is a 0.8 billion parameter Qwen3-based language model developed by andresnowak. It is a fine-tuned version of andresnowak/Qwen3-0.6B-MNLP_mcqa_model_text, specifically optimized for multiple-choice question answering (MCQA) tasks. The model was trained using the TRL library and incorporates the GRPO (Generalized Reinforcement Learning with Policy Optimization) method, as introduced in the DeepSeekMath paper.

Key Capabilities & Training

  • Reinforcement Learning Fine-tuning: The model underwent 2 epochs of RL training using a simple verifiable reward system (RLVR), where a reward of +1.0 was given for outputs containing [Letter]. and -1.0 otherwise, ensuring structured answers.
  • Multiple-Choice QA Focus: It is specifically designed to answer multiple-choice questions, with training datasets including Math QA, ScienceQA, MMLU, AI2 ARC, MedMCQA, OpenBookQA, and SciQ.
  • Structured Output: The training methodology encourages the model to produce answers in the format [Letter]. [Answer].

Performance Highlights

The model was evaluated across various MCQA benchmarks using different prompt formats. Notably, it shows strong performance in Math QA, achieving up to 35.2% accuracy (Acc) and 34.6% normalized accuracy (Acc Norm) in one evaluation setting, and 34.93% when only predicting the letter. Overall accuracies across benchmarks range from 36.27% to 47.19% depending on the prompt and evaluation type.

Good for

  • Multiple-Choice Question Answering: Particularly in STEM fields, given its training on datasets like Math QA, ScienceQA, and MMLU.
  • Structured Answer Generation: When a specific output format like [Letter]. [Answer] is required for MCQA tasks.

Finetunes
1 models