Datasets:

pgsyttch
/

dbv2_and_alfv4

base_model: N/A (Dataset Card)
datasets:
  - u-10bei/sft_alfworld_trajectory_dataset_v4
  - u-10bei/dbbench_sft_dataset_react_v2
language:
  - en
license: mit
tags:
  - dataset
  - agent
  - trajectory
  - alfworld
  - dbbench

Dataset: pgsyttch/dbv2_and_alfv4

This repository hosts a combined dataset designed for Supervised Fine-Tuning (SFT) of agent models. It merges two distinct trajectory datasets: ALFWorld (household tasks) and DBBench (database operations).

Dataset Description

This dataset is a concatenation of agent trajectories from:

ALFWorld: Provides multi-turn interaction data for household tasks.
DBBench: Offers multi-turn interaction data for database operation tasks, often in ReAct style.

The data is pre-processed into the OpenAI messages format, suitable for training conversational AI agents.

Data Fields

The dataset contains a single primary field:

messages: A list of dictionaries, where each dictionary represents a turn in a conversation. Each turn has role (e.g., system, user, assistant, tool) and content.

Usage (Example for Loading)

from datasets import load_dataset

dataset = load_dataset("pgsyttch/dbv2_and_alfv4", split="train")
print(dataset[0])

Sources & Terms

Training data originates from:

u-10bei/sft_alfworld_trajectory_dataset_v4
u-10bei/dbbench_sft_dataset_react_v2

Dataset License: mit License. This dataset is used and distributed under the terms of the mit License.