laion/explore-tis-untrunc-45-8B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 11, 2026Architecture:Transformer Featherless Exclusive Cold

The laion/explore-tis-untrunc-45-8B is an 8 billion parameter agentic RL checkpoint, finetuned from a Qwen3-8B SFT base model. This model utilizes RLOO-n advantage with truncated importance sampling (TIS) and features untruncated sampling during exploration. It is specifically optimized for agentic terminal-bench tasks, demonstrating capabilities in complex problem-solving environments.

Loading preview...

Model Overview

laion/explore-tis-untrunc-45-8B is an 8 billion parameter agentic Reinforcement Learning (RL) checkpoint, developed by laion. It is finetuned from a Qwen3-8B SFT base model using the SkyRL framework.

Key Characteristics

  • Algorithm: Employs RLOO-n advantage with seq_mean_token_sum_norm_global loss reduction and no KL loss.
  • Truncated Importance Sampling (TIS): Features enabled TIS (tis_imp_ratio_cap=2.0) to correct for logprob discrepancies between vLLM and FSDP bf16.
  • Exploration Strategy: A distinguishing feature is its untruncated sampling during the exploration phase.
  • Training Data: Trained on the DCAgent/exp_rpt_pymethods2test-large dataset.
  • Evaluation Harness: Evaluated using terminus-2 / Harbor agentic terminal-bench with 8 samples per prompt.

Performance and Selection

This specific checkpoint, global_step 45, was selected based on the highest trailing-5 Exponential Moving Average (EMA) of reward/avg_raw_reward within the genuine training region (steps <= 78). At this step, the raw reward was approximately 0.518, with a trailing-5 EMA of about 0.495.

Use Cases

This model is particularly well-suited for agentic tasks and environments that require complex problem-solving and decision-making, especially within terminal-based or code-oriented challenges, due to its specialized RL training and exploration strategy.

Finetunes
1 models