kuririrn/sft_alfworld_trajectory_dataset_v3to5_admissible_success

Dataset Card for ALFWorld SFT Trajectories with Admissible Actions (v3–v5)

Overview

This dataset contains a subset of ALFWorld-style trajectories used for supervised fine-tuning (SFT) of an agent interacting with a textual household environment.

It is constructed by merging v3, v4, and v5 of the original u-10bei/sftalfworldtrajectorydataset* series and filtering only trajectories that include admissible actions in the observation text.

Each example corresponds to one full trajectory (episode).

Source datasets:

• —u-10bei/sftalfworldtrajectorydatasetv3
• —u-10bei/sftalfworldtrajectorydatasetv4
• —u-10bei/sftalfworldtrajectorydatasetv5

Filtering criteria:

1. 1.The trajectory must contain at least one message whose content includes "Admissible actions:" (case-insensitive).
2. 2.If ISSUCCESSONLY=True during construction:
3. 3.For v5, only trajectories with trajectory_outcome == "success" are kept.
4. 4.For v3 and v4, all trajectories are retained (these versions do not explicitly annotate failure cases).

This dataset is intended for supervised fine-tuning (SFT) of ReAct-style agents with explicit admissible action constraints.

Intended Use

This dataset is designed to train or fine-tune an agent to:

• —Read a task description and environment observation.
• —Parse admissible action constraints embedded in observations.
• —Reason step-by-step ("Think:").
• —Output valid environment actions ("Act:").
• —Learn action selection consistent with the admissible action list.

It is especially suitable for:

• —ReAct-style agent fine-tuning
• —Tool-use alignment
• —Constrained action policy learning

This dataset only provides a train split. Users are expected to create their own validation/test splits if needed.

Dataset Structure

Split:

• —train: all admissible-action trajectories from v3–v5 (optionally success-filtered for v5)

Columns

messages:

• —Type: list of dictionaries
• —Format: {"role": "system" | "user" | "assistant", "content": string}

Role behavior:

• —system: Describes the overall ALFWorld environment setup.
• —user: Provides the task description and environment observation. May include a line such as: Admissible actions: [go to countertop 1, open fridge 1, ...]
• —assistant: Contains ReAct-style outputs such as: Think: ... Act: go to fridge 1

This column is directly usable in chat-style SFT pipelines.

metadata:

• —Type: structured dictionary
• —Schema aligned to v5 format
• —Includes fields such as: tasktype description roomtype difficulty numsteps hasrecoverysteps numactions numobservations numdetours trajectoryoutcome failurereason and related statistics

For v3 and v4 trajectories, fields introduced in v5 are present but may contain null values.

trajectory_outcome:

• —Type: string
• —Values: success failure unknown

Notes:

• —v5 explicitly labels success and failure.
• —v3 and v4 do not provide failure labels.
• —During construction, missing labels are set to "unknown".

dataset_version:

• —Type: string
• —Values: v3 v4 v5

Indicates which original dataset version the trajectory came from. This enables version-wise analysis and selective filtering.

Construction Procedure

1. 1.Load v3, v4, v5 original datasets.
2. 2.Add dataset_version column.
3. 3.Normalize trajectory_outcome:
4. 4.If missing, set to "unknown".
5. 5.Filter trajectories containing "admissible actions" in message content.
6. 6.Align metadata schema across versions by casting to v5 metadata structure.
7. 7.If ISSUCCESSONLY=True:
8. 8.Keep all v3 and v4 samples.
9. 9.Keep only success samples from v5.

Example Statistics

Admissible only (v3–v5 combined): 1859 trajectories. Success-only (v5 filtered): 616 trajectories. v3 + v4 + v5-success combined: approximately 1769 trajectories.

Exact numbers may vary depending on filtering flags.

Limitations

• —Only trajectories containing admissible actions are included. This biases the dataset toward constrained-environment episodes.
• —v3 and v4 lack explicit failure labels.
• —Success-only filtering affects only v5 data.
• —The dataset inherits structural assumptions from the original ALFWorld trajectory design.

License

license: other is currently used as a placeholder. Please ensure compliance with the original dataset license.

Japanese Notes (日本語メモ)

• —元データセット: u-10bei/sftalfworldtrajectorydatasetv3 u-10bei/sftalfworldtrajectorydatasetv4 u-10bei/sftalfworldtrajectorydatasetv5
• —"Admissible actions" を含む軌跡のみ抽出。
• —metadata は削除せず保持。
• —trajectory_outcome をトップレベルに追加し統一。
• —ISSUCCESSONLY=True の場合: v5 は success のみ、v3/v4 は全件保持。
• —dataset_version カラムで元バージョンを識別可能。

本データセットは、制約付き行動選択を学習する ReAct型エージェントのSFT用途を主目的としています。