laion/explore-tis-minp-40-8B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 12, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The laion/explore-tis-minp-40-8B is an 8 billion parameter language model developed by laion, derived from the GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified-maxeps-131k-fixthink base model. This model is a reinforcement learning (RL) checkpoint, specifically from an ablation study focusing on min-p sampling, and is optimized for agentic tasks using the SkyRL framework. It was trained on the DCAgent/exp_rpt_pymethods2test-large dataset, making it suitable for environments requiring robust agent behavior and decision-making.

Loading preview...

Model Overview

The laion/explore-tis-minp-40-8B is an 8 billion parameter language model developed by laion, representing a specific checkpoint from the explore-tis sampling-parameter ablation study. It is a reinforcement learning (RL) model, specifically utilizing the SkyRL framework with agentic terminal-bench, Harbor, and Daytona components.

Key Characteristics

  • Base Model: Derived from laion/GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink.
  • Training Data: Trained on the DCAgent/exp_rpt_pymethods2test-large dataset.
  • RL Algorithm: Employs RLOO-n with no KL loss, using loss_reduction=seq_mean_token_sum_norm_global and TIS (Token Importance Sampling) enabled (tis_imp_ratio_cap=2.0).
  • Sampling Strategy: Configured with temperature=1.0, min_p=0.05, top_k=-1, and top_p=1.0.
  • Checkpoint Selection: The global_step_40 checkpoint was chosen based on the best trailing-5 EMA of reward/avg_raw_reward.

Intended Use

This model is particularly suited for research and development in reinforcement learning applications, especially those involving agentic behavior and environments where min-p sampling characteristics are being explored. Its training on a specific dataset suggests its utility in tasks related to method reporting and testing within a programmatic context.

Finetunes
1 models