wckwan/ALFWorld-Qwen3-8B-Adaptive-Pivot-eval

ALFWorld Qwen3-8B Adaptive-Pivot: held-out evaluation

Held-out ALFWorld evaluation rows for the checkpoints in wckwan/Alfworld-Qwen3-8B-Adaptive-Pivot (merged_hf_actor_gs{20,40,60,80,100}; the run stopped at step 100).

Layout

global_step_N/result.json         # metrics, diagnostics, task list, config, fingerprint, runtime
global_step_N/trajectories.jsonl  # per-episode trajectories
global_step_N/eval_seed0.log      # eval log
MANIFEST.sha256                   # sha256sum -c compatible

Protocol

standalone_eval preset alfworld_qwen3, identical to the ALFWorld Qwen3-8B baseline rows:

field	value
alfworld_eval_dataset	eval_in_distribution
env_name	alfworld/AlfredTWEnv
env_seed	0
val_batch_size	128 goals
max_steps	50
max_prompt_length / max_response_length	2048 / 2048
history_length	2
temperature / top_p / top_k	0.4 / 1.0 / -1
enable_thinking	true
truncation	error
history_action_char_cap	None (uncapped)

Runtime: vLLM 0.11.0, torch 2.8.0+cu128, transformers 4.57.1, data parallel 2 (TP=1), engine seed 0, gpu_memory_utilization 0.85. Evaluated at verl-agent commit 8abb4e2d.

With 128 goals, the sampling standard deviation of a success rate is about 4.4 points. Treat differences under about 9 points as not separated.

Config fingerprint: why it differs from the baselines

	fingerprint
these rows (native)	8bf7e38499b77e241a766319deed699770df753a691d71eda415132948971c80
baseline rows / legacy-equivalent	4b8cad2e9b47086439ab8f3b65fad564769b958efcf67280ea40547a0e1e7c68

The difference is a schema change, not a protocol change. Commit a668c6b6 (2026-09-10) added the config field history_action_char_cap, and it is None here. The baseline rows predate the field, so the key is absent from their configs. Removing that single key from this config reproduces 4b8cad2e… exactly, which shows no other field differs. With the cap at None, history_actions in agent_system/environments/base.py returns the action list unchanged, so prompts, rewards and action validity are identical to the historical behaviour.

To aggregate these rows with the baselines, pass --allow-mixed to standalone_eval.sweep (aggregate mode). That is safe for exactly the reason above.