laion/ablation-pymethods2test-shaped-45-8B

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 8, 2026Architecture:Transformer Featherless Exclusive Cold

The laion/ablation-pymethods2test-shaped-45-8B is an 8 billion parameter language model, based on a Qwen3-8B SFT model, fine-tuned using Reinforcement Learning (RL) with a shaped reward function. Developed by laion, this model is part of an ablation study focusing on reward shaping, specifically optimizing for the fraction of passing tests rather than a binary all-tests-pass reward. It is designed for tasks involving code generation and testing, aiming to improve performance by rewarding partial correctness.

Loading preview...

Model Overview

The laion/ablation-pymethods2test-shaped-45-8B is an 8 billion parameter model derived from a Qwen3-8B SFT base model (laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink). This model is a specific checkpoint from an RL-based ablation study, focusing on the impact of shaped rewards during training.

Key Differentiators

  • Shaped Reward Function: Unlike models trained with a binary pass/fail reward, this model was optimized using a "shaped pass-ratio" reward. This means it was rewarded based on the fraction of tests passing, encouraging incremental improvements in code correctness.
  • RL Training: The model was fine-tuned using SkyRL GRPO, a Reinforcement Learning algorithm, over 80 training steps on 14x GH200 nodes.
  • Training Dataset: It utilized the DCAgent/exp_rpt_pymethods2test-large dataset, which likely contains report-style data related to Python methods and tests.
  • Checkpoint Selection: The global_step_45 checkpoint was selected based on the Exponential Moving Average (EMA) of the raw reward, indicating a robust performance point during training.

Use Cases

This model is particularly relevant for research in:

  • Code Generation: Especially for scenarios where partial correctness or incremental improvements in test pass rates are valuable.
  • Reinforcement Learning for Language Models: Demonstrating the effects of different reward shaping strategies.
  • Automated Testing and Code Repair: Where understanding and improving code based on test outcomes is critical.

Finetunes
1 models