laion/a3-rl-laion_exp_rpt_methods2test-large-v2

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 6, 2026Architecture:Transformer Featherless Exclusive Cold

The laion/a3-rl-laion_exp_rpt_methods2test-large-v2 is an 8 billion parameter model, partially fine-tuned using Reinforcement Learning (RLOO-n, token_mean) on a Qwen3-8B base. It was trained on the methods2test-large-v2 agentic task set, focusing on improving performance in agentic tasks. This model represents a partial checkpoint from an experiment series that was concluded as uninformative, but showed healthy training progress at cancellation.

Loading preview...

Overview

This model, a3-rl-laion_exp_rpt_methods2test-large-v2, is an 8 billion parameter language model derived from a Qwen3-8B base. It underwent Reinforcement Learning (RL) fine-tuning using the RLOO-n algorithm with token_mean on the methods2test-large-v2 agentic task set. The training was conducted via SkyRL terminal-bench / terminus-2.

Key Characteristics

  • Base Model: Qwen3-8B (specifically laion/GLM-4_7-swesmith...-fixthink).
  • Fine-tuning Method: Reinforcement Learning (RLOO-n, token_mean) with eps_clip 0.2/0.05 and lr 8e-6.
  • Training Status: This is a partial checkpoint (global_step_50) as the a3 experiment series was cancelled mid-run due to being deemed uninformative. Despite cancellation, training metrics showed healthy progress, with avg_raw_reward improving from ~0.43 to ~0.70 and pass@8 from ~0.81 to ~0.88, while entropy remained stable.

Potential Use Cases

Given its RL fine-tuning on an agentic task set, this model could be explored for:

  • Agentic Task Development: As a base for further experimentation or fine-tuning on specific agentic workflows.
  • RL Research: Investigating the effects of partial RL training on Qwen3-8B for agentic capabilities.
  • Benchmarking: Evaluating its performance on tasks similar to methods2test-large-v2 to understand the impact of the specific RL approach.

Finetunes
1 models