Jazhyc/Llama-3.1-8B-aims-grpo

TEXT GENERATIONPricing:Input $0.2 / Cached $0.028 / Output $0.32Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 23, 2026License:llama3.1Architecture:Transformer Featherless Exclusive Cold

Jazhyc/Llama-3.1-8B-aims-grpo is an 8 billion parameter, Llama-3.1-based model fine-tuned by Jazhyc for intent-aware, reasoning-based safety classification. It generates a reasoning trace, explicit intent, and a binary harm label for user prompts. Trained with Group Relative Policy Optimization (GRPO) and an intent-faithfulness reward, it achieves an average F1 score of 0.836 across five external safety benchmarks, making it suitable for prompt-level moderation research.

Loading preview...

Overview

Jazhyc/Llama-3.1-8B-aims-grpo is a full fine-tuned model derived from meta-llama/Llama-3.1-8B-Instruct. It functions as an intent-aware, reasoning-based safety classifier, designed to analyze user prompts and produce a detailed <reasoning> trace, an explicit intent, and a binary harm label (harmful/safe).

Key Differentiators

  • Intent-Faithfulness Reward: This model is trained using GRPO (Group Relative Policy Optimization) with a unique reward mechanism that explicitly verifies intent faithfulness against human annotations from the AIMS dataset. This approach yields superior performance compared to label-only reward systems.
  • Strong Performance: It achieves the best average F1 score of 0.836 across five external safety benchmarks (WildGuardTest, XSTest, AEGIS 2.0, ToxicChat, OpenAI Mod), outperforming other Llama-3.1-8B variants and ablation models.
  • Structured Output: For each classification, it provides a structured output including a reasoning trace and a clear intent statement, enhancing transparency and interpretability.

Intended Use Cases

  • Research on Safety Classification: Ideal for academic and industrial research into intent-aware and reasoning-based safety classification systems.
  • Prompt-Level Moderation: Can be used as a classifier for moderating user prompts, identifying potentially harmful content based on underlying intent.

Limitations

  • English-only: The model is trained exclusively on English data.
  • Prompt-focused: It classifies the prompt itself, not model responses.
  • Slower Classification: Due to generating a reasoning trace, it is slower than direct-classification variants.
  • Not a Sole Authority: Outputs should not be treated as the sole authority for high-stakes moderation decisions.

Finetunes
1 models