---
language:
- en
license: apache-2.0
task_categories:
- text-generation
tags:
- alfworld
- dbbench
- agentbench
- agent
- multi-turn
- sft
- deep-thinking
size_categories:
- 1K<n<10K
---

# Deepthinking ALFWorld & DBBench Spider v2

AgentBench 評価の 2 タスク（ALFWorld / DBBench）を統合した **マルチタスク SFT 訓練データセット**。
フォーマット検査・フィルタリング済みの **7,779 件**を、サイズ比率に基づく等間隔インターリーブで結合。

---

## Dataset Summary

| Metric | Value |
|--------|-------|
| **Total rows** | **7,779** |
| ALFWorld | 4,884 (62.8%) |
| DBBench | 2,895 (37.2%) |
| Avg messages per item | 18.3 |
| Columns | `messages` |
| Interleave method | 比率ベース等間隔マージ |

---

## Source Datasets

| Source | Rows | Description |
|--------|------|-------------|
| [mark-22/Deepthinking-sft_alfworld_final1](https://huggingface.co/datasets/mark-22/Deepthinking-sft_alfworld_final1) | 4,884 | ALFWorld タスク。GPT-OSS-120B で最初の THOUGHT に状況分析・常識推論・ステップ分解を挿入したデータ拡張版 |
| [mark-22/dbbench_spider_v4_mergeddata_final1](https://huggingface.co/datasets/mark-22/dbbench_spider_v4_mergeddata_final1) | 2,897 → 2,895 | DBBench タスク。Spider 生成データ (1,697) + u-10bei クレンジング済みデータ (1,200) の統合。フォーマット検査で 2 件除外 |

---

## Merge Pipeline

```
mark-22/Deepthinking-sft_alfworld_final1 (4,884)
  ↓ フォーマット検査: THOUGHT + ACTION パターン必須
  ↓ 除外: 0 件
  ↓ シャッフル (seed=42)
  ↓
  ├─────────────────────────────────────────────┐
  │  等間隔インターリーブ (サイズ比率ベース)      │
  ├─────────────────────────────────────────────┘
  ↓
mark-22/dbbench_spider_v4_mergeddata_final1 (2,897)
  ↓ フォーマット検査: Action: Operation/Answer 必須
  ↓ 除外: 2 件（フォーマット違反）
  ↓ シャッフル (seed=42)
  ↓
  → 7,779 件（ALF:DB ≈ 63:37）
```

### Formatting Filters

#### ALFWorld
- 全 assistant メッセージに `THOUGHT:` と `ACTION:` の両方が含まれること
- `THOUGHT:` が `ACTION:` より前に出現すること

#### DBBench
- `"Ok."` 等の承認応答はスキップ
- それ以外の assistant メッセージに `Action: Operation` または `Action: Answer` が含まれること
- 少なくとも 1 回は Action が実行されていること

### Interleave Strategy

単純な連結ではなく、**比率ベースの等間隔マージ**を採用:

- 各データセットを独立シャッフル後、正規化位置 `(i + 0.5) / N` を付与
- 全アイテムを正規化位置でソートし、ALFWorld と DBBench が均等に混ざるよう配置
- 訓練時にタスクが偏らず、安定した学習を実現

```
結果のインターリーブパターン（先頭10件）:
  [0] ALF  [1] DB  [2] ALF  [3] ALF  [4] DB
  [5] ALF  [6] DB  [7] ALF  [8] ALF  [9] DB
```

---

## Conversation Structures

### ALFWorld items

```
[system]    Interact with a household to solve a task...
[user]      You arrive in a room. You can see...
[assistant] THOUGHT: {AI-generated analysis} + {original thought}
            ACTION: {action command}
[user]      Environment response
...
```

### DBBench items

```
[user]      I will ask you a question, then you should help me operate a MySQL database...
[assistant] Ok.
[user]      Question text
[assistant] Thought + Action: Operation ```sql ... ```
[user]      MySQL raw output
...
[assistant] ... Action: Answer  Final Answer: ["..."]
```

---

## Usage

```python
from datasets import load_dataset

ds = load_dataset("mark-22/Deepthinking-alfworld_and_dbbench_spider_v2", split="train")

for item in ds:
    messages = item["messages"]
    # Detect task type by first message role:
    #   system -> ALFWorld
    #   user   -> DBBench
    task_type = "alfworld" if messages[0]["role"] == "system" else "dbbench"
```

---

## Merge Code

<details>
<summary>結合スクリプト全文（クリックで展開）</summary>

```python
import os, re
from datasets import load_dataset, Dataset, Features, Value
from huggingface_hub import login

ds_alf = load_dataset("mark-22/Deepthinking-sft_alfworld_final1", split="train")
ds_db  = load_dataset("mark-22/dbbench_spider_v4_mergeddata_final1", split="train")

# Unify columns to "messages" only
ds_alf = ds_alf.select_columns(["messages"])
ds_db  = ds_db.select_columns(["messages"])

unified_features = Features({
    "messages": [{"role": Value("string"), "content": Value("string")}]
})

def fix_message_order(example):
    return {"messages": [
        {"role": msg["role"], "content": msg["content"]}
        for msg in example["messages"]
    ]}

ds_alf = ds_alf.map(fix_message_order, features=unified_features)
ds_db  = ds_db.map(fix_message_order, features=unified_features)

# Format validation
def is_valid_alfworld(example):
    for msg in example["messages"]:
        if msg["role"] == "assistant":
            content = msg["content"]
            t = re.search(r"THOUGHT\s*:", content, re.IGNORECASE)
            a = re.search(r"ACTION\s*:", content, re.IGNORECASE)
            if not t or not a or t.start() > a.start():
                return False
    return True

def is_valid_dbbench(example):
    has_action = False
    for msg in example["messages"]:
        if msg["role"] == "assistant":
            content = msg["content"].strip()
            if content.lower() in ["ok", "ok.", "understood", "understood.", "yes", "yes."]:
                continue
            if not re.search(r"Action:\s*(Operation|Answer)", content, re.IGNORECASE):
                return False
            has_action = True
    return has_action

ds_alf = ds_alf.filter(is_valid_alfworld)
ds_db  = ds_db.filter(is_valid_dbbench)

# Independent shuffle + ratio-based interleave
ds_alf = ds_alf.shuffle(seed=42)
ds_db  = ds_db.shuffle(seed=42)

alf_items = [((i + 0.5) / len(ds_alf), item) for i, item in enumerate(ds_alf)]
db_items  = [((i + 0.5) / len(ds_db),  item) for i, item in enumerate(ds_db)]
combined  = sorted(alf_items + db_items, key=lambda x: x[0])
merged_ds = Dataset.from_list(
    [item for _, item in combined], features=unified_features
)

merged_ds.push_to_hub("mark-22/Deepthinking-alfworld_and_dbbench_spider_v2")
```

</details>

---

## License

Apache 2.0