laion/a3-rl-DCAgent_exp_rpt_e2egit-large_global_step_15

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 28, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

laion/a3-rl-DCAgent_exp_rpt_e2egit-large_global_step_15 is an 8 billion parameter reinforcement learning (RL) model developed by laion, based on a Qwen3-8B SFT base model. This checkpoint, global_step_15, was selected for its optimized reward performance during a SkyRL terminal-bench RL run. It is specifically fine-tuned for tasks related to the DCAgent/exp_rpt_e2egit-large dataset, making it suitable for environments requiring robust RL policy execution.

Loading preview...

Model Overview

This model, laion/a3-rl-DCAgent_exp_rpt_e2egit-large_global_step_15, is an 8 billion parameter reinforcement learning (RL) agent. It is built upon a laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink base model, which is a Qwen3-8B SFT variant. The model was developed through a SkyRL terminal-bench RL run, with global_step_15 identified as the optimal checkpoint based on an EMA (reward/avg_raw_reward, 5-step) metric, achieving an EMA of 0.8639 and a reward of 0.8418.

Key Training Details

  • Base Model: Qwen3-8B SFT (laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink)
  • Task Dataset: /e/scratch/jureap59/feuer1/tasks/exp_rpt_e2egit-large (DCAgent/exp_rpt_e2egit-large)
  • Algorithm: RLOO-n with specific hyperparameters (eps_clip_low=0.2, eps_clip_high=0.05, kl=0, lr=8e-6).
  • Trainer: SkyRL FSDP2, utilizing 14 nodes with 4 GPUs per node, a training batch size of 64, and 48 vLLM engines.
  • Total Steps Trained: 80 steps, though the job was cancelled after a performance collapse beyond step 65.

Training Traces and Logs

Companion training-time Daytona/Harbor rollouts are available as a dataset: open-athena/a3-rl-DCAgent_exp_rpt_e2egit-large. This dataset contains the last episode of each trial, representing the rollouts the policy was trained on. Training logs, including metrics.csv, vllm_metrics.csv, trial_stats.csv, report.md, and reward_plot.png, are provided in the training_logs/ directory, alongside raw trainer_log.jsonl and *.out files for archival purposes.

Good for

  • Applications requiring an RL policy fine-tuned on the DCAgent/exp_rpt_e2egit-large dataset.
  • Research and development in reinforcement learning, particularly for analyzing RL training dynamics and checkpoint selection based on EMA metrics.

Finetunes
1 models