psymon/mistral-7b-mio-arc-fp16

TEXT GENERATIONConcurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:4kTool Calling:SupportedPublished:Jul 15, 2026License:apache-2.0-mixed-data-termsArchitecture:Transformer Open Weights Featherless Exclusive Cold

psymon/mistral-7b-mio-arc-fp16 is a 7 billion parameter Mistral-7B-v0.1 derivative, fine-tuned by psymon, specifically optimized for science and commonsense multiple-choice reasoning tasks. This model achieved 76.54% acc_norm on the ARC-Challenge test split, making it highly specialized for short English science-QA completions. It was developed using a two-stage training process involving QLoRA SFT and Mutual Information Optimization (MIO) preference fine-tuning.

Loading preview...

Overview

psymon/mistral-7b-mio-arc-fp16 is a 7 billion parameter model derived from mistralai/Mistral-7B-v0.1, specifically fine-tuned for science and commonsense multiple-choice reasoning. It was developed through a two-stage process: initial QLoRA Supervised Fine-Tuning (SFT) on 27,294 public science-QA examples, followed by Mutual Information Optimization (MIO) preference fine-tuning on 47,087 train-only pairs. This MIO stage used hard-negative selection, where rejected answers were the incorrect choices the SFT model found most plausible.

Key Capabilities

  • High Accuracy on ARC-Challenge: Achieved 76.54% acc_norm and 74.74% raw accuracy on the 1,172-example ARC-Challenge test split with 25-shot evaluation.
  • Domain-Specialized Reasoning: Optimized for short English science-QA completions, demonstrating significant improvement in this specific domain.
  • Preference Fine-Tuning: Utilizes a unique MIO preference fine-tuning approach with model-scored hard negatives to enhance reasoning capabilities.

Limitations and Considerations

  • Domain Specialization: This is a domain-specialized completion model, not a general chat or instruction-following model. It is not intended for safety-critical use or factual deployment without verification.
  • Capability Regressions: Audits showed regressions in MMLU Humanities, MMLU Social Sciences, and WikiText perplexity, indicating a trade-off in general capabilities for specialized performance.
  • Prompt Format: Expects a specific prompt format for questions, focusing on answer-text completions rather than answer-position labels.

Intended Use Cases

  • Evaluating science and commonsense multiple-choice reasoning.
  • Research into model-scored hard-negative selection and preference optimization for short-answer QA.
  • Reproducing the reported ARC-Challenge evaluation results.

Finetunes
1 models