YYYYYYibo/alfworld-expert-prefix-rollouts

ALFWorld Expert-Prefix Rollout Landscape

This dataset measures how a frozen language-model actor's probability of solving an ALFWorld task changes after replaying different-length prefixes of a successful expert trajectory.

The collection contains all 3,553 ALFWorld training tasks from the Agent-G2 SFT data. Eight independent actor rollouts were sampled from the initial state for every task. For the 2,307 low-signal tasks with at most one root success, eight additional continuations were sampled after every possible non-full expert prefix, giving 17,621 prefix points and 140,968 prefix continuations.

The actor was Qwen2.5-7B-Instruct with temperature 0.8, top-p 0.95, and a total budget of 20 environment actions. An expert prefix of depth d leaves 20-d actor actions. Exact collection metadata is retained in the Parquet records and in collection_metadata.json.

Headline results

Across the 2,307 low-signal tasks, 1,473 (63.85%) have at least one valid prefix position with 3--5 successes out of 8, and 733 (31.77%) have a position with exactly 4/8 successes. The most learnable position occurs at a mean normalized expert depth of 0.469. See analysis/final_summary.json for complete per-depth and per-task-type statistics.

Configurations

• prefix_points (default): one row per task/prefix depth, including the expert prefix replay, next expert action, 8-bit outcome vector, success estimate, and learnability score 4p(1-p).
• prefix_rollouts: one row per actor continuation from a prefix point. Join on prefix_id. rollout_json losslessly retains every generated step, prompt, observation, admissible-action list, generation metadata, and reward.
• root_rollouts: one row per rollout from the initial task state.
• tasks: root-level aggregate metadata for all tasks.
• experts: the complete successful expert think/action trajectories from the upstream Agent-G2 dataset.

Nested finalized traces are stored as compact JSON strings to guarantee that heterogeneous step metadata is retained exactly. Parse rollout_json, prefix_metadata_json, or root_metadata_json with json.loads.

import json
from datasets import load_dataset

points = load_dataset(
    "YYYYYYibo/alfworld-expert-prefix-rollouts",
    "prefix_points",
    split="train",
)
rollouts = load_dataset(
    "YYYYYYibo/alfworld-expert-prefix-rollouts",
    "prefix_rollouts",
    split="train",
    streaming=True,
)

first = next(iter(rollouts))
steps = json.loads(first["rollout_json"])["steps"]

Important semantics

• depth=0 is a fresh set of eight continuations from the initial state; it is distinct from the eight screening samples in root_rollouts.
• normalized_depth = depth / expert_length.
• p_hat = num_successes / num_continuations is based on eight samples and has substantial binomial uncertainty.
• prefix_terminal=true marks 11 positions where replaying the prefix had already terminated the environment. The actor was skipped for those rows; exclude them when studying handoff-position learnability.
• The landscape was collected only for root-low-signal tasks (0/8 or 1/8).

Provenance and license

Expert trajectories come from xiamoent/Agent-G2-ALFWorld-Webshop-sft-data, which is released under Apache-2.0. Users must also comply with the licenses and terms of ALFWorld and the underlying TextWorld data.

The per-step event logs used as crash-recovery journals are intentionally not included because they duplicate the finalized rollout records.

Limitations

This is a stochastic snapshot from one actor model and one sampling setup, not a ground-truth task-difficulty label. The generated reasoning may contain errors, and success estimates based on eight samples are noisy.