sol-max-v2.h013.pi-agent-sft-v10.step_125

sol-max-v2.h013.pi-agent-sft-v10.step_125

AgentPTB sweep checkpoint. Cell sol-max-v2 — Codex / gpt-5.6-sol @ effort max.

Reading the eos field

248046 is <|im_end|>, the token the Qwen3.5 chat template ends every assistant turn with. Checkpoints missing it do not stop at end-of-turn and overrun the context window, so their eval numbers are a floor, not a measurement — compare them only against other checkpoints with the same eos status, or re-package before evaluating.

Cell note: redo of the sol@max cell from hour 0 on tb-1; ran the full 100 h. The arm submitted an h7 checkpoint over 75 h of later training.

Serving

Qwen/Qwen3.5-9B-Base is Qwen3_5ForConditionalGeneration — a vision architecture, and the vision tower is present in these weights. prime-rl does not export preprocessor_config.json, so vLLM must be told the model is text-only or it fails to load:

--limit-mm-per-prompt '{"image": 0, "video": 0}'

Mapping back to the figures

The repo id is {cell}.h{HHH}.{family}.{step}, where hHHH is the hour of the 100-hour run at which this checkpoint was written — the same x-axis the sweep figures use for eval panels (t_h). So a checkpoint drops onto the performance-over-time curve directly, and sorting repo ids within a cell sorts them chronologically.

hHHH is rounded down to whole hours for sortability; the exact value is the hours into run row above, and in agentic-ptb/INDEX.