---
dataset_info:
  features:
  - name: messages
    list:
    - name: role
      dtype: string
    - name: content
      dtype: string
  - name: metadata
    struct:
    - name: task_type
      dtype: string
    - name: description
      dtype: string
    - name: room_type
      dtype: string
    - name: difficulty
      dtype: string
    - name: num_steps
      dtype: int64
    - name: has_recovery_steps
      dtype: bool
    - name: has_close_open
      dtype: bool
    - name: has_admissible
      dtype: bool
    - name: num_actions
      dtype: int64
    - name: num_observations
      dtype: int64
    - name: num_detours
      dtype: int64
    - name: num_recovery_steps
      dtype: int64
    - name: has_open_action
      dtype: bool
    - name: subgoals
      sequence: string
    - name: total_chars
      dtype: int64
    - name: trajectory_outcome
      dtype: string
    - name: failure_reason
      dtype: string
    - name: type
      dtype: string
    - name: table_name
      dtype: string
    - name: sql
      dtype: string
    - name: label
      sequence: string
    - name: wrong_sql
      dtype: string
    - name: wrong_val
      dtype: string
    - name: correct_val
      dtype: string
    - name: where_col
      dtype: string
    - name: error_col
      dtype: string
    - name: error_reason
      dtype: string
  - name: id
    dtype: string
  splits:
  - name: train
    num_bytes: 7988195.684038021
    num_examples: 5584
  download_size: 1566110
  dataset_size: 7988195.684038021
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# mixed-agent-dataset-v6_v2

This dataset is an improved mixed training dataset for AgentBench tasks.

It combines ALFWorld and DBBench with task-specific normalization.

## Source Datasets

- https://huggingface.co/datasets/u-10bei/sft_alfworld_trajectory_dataset_v5
- https://huggingface.co/datasets/u-10bei/dbbench_sft_dataset_react_v4

## Preprocessing

To stabilize multi-task learning, outputs were normalized.

### ALFWorld

Assistant outputs were converted to **action-only format**.

Example:

Before

Think: I should open the drawer  
Act: open drawer 1

After

Act: open drawer 1

### DBBench

Assistant responses were normalized to **SQL-only format**.

Example

User: database question  
Assistant: SQL query

## Data Cleaning

Duplicate samples were removed using message-level hashing.

messages → canonical string → SHA1 → dedup

## Dataset Purpose

The goal is to improve training stability for agent-style tasks.

Key improvements:

- consistent output formats
- reduced noisy reasoning tokens
- cleaner supervision signal

## Dataset Size

- total samples: ~5k
- split: train

## Intended Use

Supervised fine-tuning (SFT) for LLM agent models in the AgentBench competition.