longtermrisk/Llama-3.1-8B-school-of-reward-hacks-sft-seed5

TEXT GENERATIONPricing:Input $0.37 / Cached $0.074 / Output $0.38Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Aug 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The longtermrisk/Llama-3.1-8B-school-of-reward-hacks-sft-seed5 is an 8 billion parameter Llama-3.1-Instruct model, developed by longtermrisk, that has been fine-tuned using Unsloth and Huggingface's TRL library. This model is optimized for specific reward hacking scenarios, offering specialized performance for tasks related to understanding and mitigating reward-based vulnerabilities. Its fine-tuning process emphasizes efficiency, having been trained twice as fast as standard methods.

Loading preview...

Model Overview

This model, developed by longtermrisk, is a fine-tuned variant of the 8 billion parameter Llama-3.1-Instruct architecture. It leverages the efficiency of Unsloth and Huggingface's TRL library for its training process, enabling it to be trained twice as fast as conventional methods.

Key Characteristics

  • Base Model: Fine-tuned from unsloth/Meta-Llama-3.1-8B-Instruct.
  • Training Efficiency: Utilizes Unsloth for accelerated training, achieving 2x faster fine-tuning.
  • Specialized Fine-tuning: The model's name, "school-of-reward-hacks-sft-seed5," indicates a specific focus on reward hacking scenarios, suggesting it has been trained to understand or address behaviors related to exploiting reward functions.

Potential Use Cases

  • Research into Reward Hacking: Ideal for researchers studying the dynamics of reward functions and potential vulnerabilities in AI systems.
  • AI Safety and Alignment: Can be used to explore and develop strategies for mitigating undesirable behaviors arising from reward maximization.
  • Specialized Language Understanding: May offer unique insights into text related to game theory, reinforcement learning, and adversarial AI contexts where reward structures are critical.