agentic-ptb/sol-max.h015.baseline-bench.step_150

Overview

This model, sol-max.h015.baseline-bench.step_150, is an intermediate checkpoint from the AgentPTB sweep, specifically from the sol-max cell. It is a 9 billion parameter model built upon the Qwen/Qwen3.5-9B-Base architecture, featuring a 32,768 token context length. The training process was driven by a combination of Codex and gpt-5.6-sol with maximum reasoning effort.

Key Characteristics

• Base Model: Qwen/Qwen3.5-9B-Base
• Parameter Count: 9 billion
• Context Length: 32,768 tokens
• Training Driver: Codex / gpt-5.6-sol at max effort
• Checkpoint Timing: Captured at 16.05 hours into a 100-hour run.
• Correct eos_token_id: Crucially, this checkpoint correctly implements eos_token_id [248044, 248046] (including <|im_end|>), which is vital for accurate evaluation as it prevents context window overruns.

Importance of eos_token_id

The correct eos_token_id configuration is a significant differentiator. Checkpoints lacking this proper setup often yield evaluation numbers that are a floor, not a true measurement, due to overrunning the context window. This model's correct configuration ensures its evaluation metrics are reliable and directly comparable against other models with similar eos status.

Use Cases

This model is suitable for tasks requiring a robust 9B parameter model with a substantial context window, particularly where accurate end-of-turn behavior is critical for reliable performance and evaluation. Its origin from an agentic sweep suggests potential for tasks involving complex reasoning or multi-turn interactions.