opus-high-v2.h009.sft_sst.step_60

AgentPTB sweep checkpoint. Cell opus-high-v2 — Claude Code / claude-opus-5 @ effort high.

field	value
plot cell	opus-high-v2
driver	Claude Code / claude-opus-5
reasoning effort	high
run boot (UTC)	2026-08-19T19:35:52Z
role	intermediate
hours into run	h9.05 of 100
checkpoint path in run	ckpt/sft-sst/weights/step_60
shards	4
size	18.8 GB
base model	Qwen/Qwen3.5-9B-Base
eos_token_id	[248044, 248046] ✅ correct

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.

Cell note: rerun of the opus@high cell. The arm measured five SFT runs and found all five regressed, so it submitted the base model's tensors, unchanged — its entire reported gain comes from the pi_plus harness and an eos fix, not from training. These six checkpoints are those rejected SFT runs; they are the cell's training trajectory, not its submission.

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:

--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. Exact value is the hours into run row above, and in agentic-ptb/INDEX.