wckwan/WebShop-Olmo3-7B-SEED-eval

WebShop OLMo-3-7B-Instruct SEED: held-out evaluation

Held-out WebShop evaluation rows for the checkpoints in wckwan/WebShop-Olmo3-7B-SEED (merged_hf_actor_gs20 through gs400, every 20 steps).

Layout

global_step_N/result.json         # metrics, diagnostics, goal set, 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 webshop_olmo3 (the Instruct preset, not webshop_olmo3_think), matching the training caps of every WebShop OLMo-3 arm:

field	value
env_name	Webshop (small catalog, synthetic goals)
env_seed	0
val_batch_size	500 goals, 4 shards of 125 (default --val-shard-size)
max_steps	15
max_prompt_length / max_response_length	4096 / 2048
history_length	2
temperature / top_p / top_k	0.4 / 1.0 / -1
truncation	error
history_action_char_cap	None (uncapped)

Action parser: verl-agent commit 8abb4e2d, which includes 451b4969, so the action is read after the final </think>. This is the parser the SEED arm trained under.

Runtime: vLLM 0.11.0, torch 2.8.0, transformers 4.57.1, data parallel 2 (TP=1), engine seed 0, gpu_memory_utilization 0.85.

Goal set

Every row evaluates the canonical 500-goal sequence shared by all WebShop series, so rows pair goal for goal with the baselines:

	sha256 of json.dumps(...)
tasks.goal_indices	3c6e0c423c7234e8a59d9a8f1193f9e8be9d3c8d3874647c1ad7e54635dba05e
tasks.shards	0bc99e50663610731defb44055ace1d15f490037d6a11eea4b70d3752e042b27

scripts/check_result_webshop.py enforces both hashes, the 4x125 sharding and the 4096/2048 caps before a row is published.

Config fingerprint

	fingerprint
these rows (native)	6448dfad6bbdf8435a3bbd73b9822b19de99b45419d8921622329869f1065c38
GiGPO baseline / legacy-equivalent	060ad196f1509810037101e0b3d3c3bed1cac46ee0ac26a4e9190634367a8070

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, which leaves the history buffer unchanged. Removing that single key reproduces the GiGPO baseline hash exactly. Aggregate against baselines with --allow-mixed.

The GRPO baseline rows fingerprinted 9c807b420392 were evaluated at max_response_length=512, a quarter of the trained cap, so they are not comparable. They have been re-evaluated at 2048 in wckwan/WebShop-Olmo3-7B-GRPO-eval-2048; compare SEED against those rows, not the 512-cap ones.