KhanCold/llama3-8b-spader

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 17, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

KhanCold/llama3-8b-spader is an 8 billion parameter Llama-3.1-8B-Instruct model fine-tuned using the SPADER reinforcement learning framework. Developed by Qiming Shi et al., SPADER introduces Step-wise Peer Advantage and a Diversity-Aware Exploration Reward. This model is specifically designed for long-horizon tool-use agents in Multi-Answer Question Answering tasks, excelling at discovering diverse entities and assigning credit effectively.

Loading preview...

Model Overview

KhanCold/llama3-8b-spader is an 8 billion parameter Llama-3.1-8B-Instruct model that has been fine-tuned using the novel SPADER (Step-wise Peer Advantage with Diversity-Aware Exploration Rewards) reinforcement learning framework. Developed by Qiming Shi, Zhaolu Kang, Yunfan Zhou, Di Weng, and Yingcai Wu, this model is specifically optimized for Multi-Answer Question Answering (QA) tasks, particularly those requiring long-horizon tool use.

Key Innovations

The SPADER framework introduces two core components:

  • Step-wise Peer Advantage (SPA): This is a critic-free mechanism for assigning credit at the step-level. It aligns parallel trajectories by decision step and estimates advantages based on peer returns, improving the learning process for complex, multi-step tasks.
  • Diversity-Aware Exploration Reward: This reward system is designed to promote the discovery of long-tail entities. It achieves this by upweighting rare findings and downweighting redundant ones, encouraging the model to explore a wider range of potential answers and information.

Use Cases

This model is particularly well-suited for applications involving:

  • Multi-Answer Question Answering: Excelling in scenarios where multiple correct answers or diverse information points are required.
  • Long-Horizon Tool-Use Agents: Its training with SPADER makes it effective for tasks that involve sequential decision-making and interaction with external tools over an extended process.
  • Enhanced Exploration: The diversity-aware reward mechanism helps in uncovering less obvious or more varied information, which can be crucial in comprehensive QA systems.

For more technical details, refer to the associated paper: SPADER: Step-wise Peer Advantage with Diversity-Aware Exploration Rewards for Multi-Answer Question Answering.

Finetunes
1 models