hyunseoki/verl-math-transfer-llama31-8b-to-llama32-3b-pool7to1

TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:Mar 29, 2026Architecture:Transformer Cold

The hyunseoki/verl-math-transfer-llama31-8b-to-llama32-3b-pool7to1 model is an 8 billion parameter LlamaForCausalLM architecture, developed by hyunseoki, specifically designed for math transfer experiments. This model focuses on transferring mathematical capabilities from a Llama 3.1 8B base to a Llama 3.2 3B configuration. It is optimized for tasks requiring mathematical reasoning and problem-solving, making it suitable for research in model capability transfer.

Loading preview...

Model Overview

This model, developed by hyunseoki, is an experimental transfer of mathematical capabilities from a Llama 3.1 8B model to a Llama 3.2 3B configuration, utilizing the verl framework. It is built on the LlamaForCausalLM architecture and comprises 8 billion parameters.

Key Capabilities

  • Mathematical Transfer Learning: Specifically trained to transfer and enhance mathematical reasoning abilities between different Llama model versions.
  • Experimental Checkpoints: Provides multiple step revisions (e.g., step-010 to step-080), allowing users to access different stages of the training process.
  • Hugging Face Compatibility: Exported into Hugging Face safetensors format, ensuring easy integration with the transformers library.

Use Cases

  • Mathematical Research: Ideal for researchers exploring transfer learning techniques in mathematical domains.
  • Model Analysis: Useful for analyzing the evolution of mathematical capabilities across different training steps and model architectures.
  • Custom Math-focused Applications: Can serve as a base for fine-tuning or developing applications that require strong mathematical understanding and problem-solving.