Add dataset card
Browse files
README.md
ADDED
|
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- en
|
| 4 |
+
license: apache-2.0
|
| 5 |
+
size_categories:
|
| 6 |
+
- 1K<n<10K
|
| 7 |
+
task_categories:
|
| 8 |
+
- text-generation
|
| 9 |
+
pretty_name: Agentic DPO Tool-Use Pairs (3K)
|
| 10 |
+
tags:
|
| 11 |
+
- dpo
|
| 12 |
+
- preference-pairs
|
| 13 |
+
- tool-use
|
| 14 |
+
- agentic
|
| 15 |
+
- rlhf
|
| 16 |
+
- function-calling
|
| 17 |
+
- alignment
|
| 18 |
+
- synthetic
|
| 19 |
+
configs:
|
| 20 |
+
- config_name: default
|
| 21 |
+
data_files:
|
| 22 |
+
- split: train
|
| 23 |
+
path: agentic-dpo-tool-use-3k.jsonl
|
| 24 |
+
---
|
| 25 |
+
|
| 26 |
+
# Agentic DPO Tool-Use Pairs (3K)
|
| 27 |
+
|
| 28 |
+
Synthetic DPO preference pairs for training LLMs to use tools correctly in agentic settings.
|
| 29 |
+
|
| 30 |
+
## Dataset Description
|
| 31 |
+
|
| 32 |
+
**3,000 preference pairs** covering 7 tool categories:
|
| 33 |
+
- `web_search` — real-time web search
|
| 34 |
+
- `calculator` — mathematical expression evaluation
|
| 35 |
+
- `weather_api` — current weather retrieval
|
| 36 |
+
- `code_interpreter` — Python code execution
|
| 37 |
+
- `database_query` — SQL database queries
|
| 38 |
+
- `stock_price` — financial data lookup
|
| 39 |
+
- `translate` — multilingual translation
|
| 40 |
+
|
| 41 |
+
Each example contains:
|
| 42 |
+
- `system`: system prompt with tool definitions
|
| 43 |
+
- `prompt`: user query requiring tool use
|
| 44 |
+
- `chosen`: correct response that uses the appropriate tool and interprets results
|
| 45 |
+
- `rejected`: flawed response (refuses tool use / hallucinates / ignores tool / claims no access)
|
| 46 |
+
- `metadata`: category and contextual info
|
| 47 |
+
|
| 48 |
+
## Format
|
| 49 |
+
|
| 50 |
+
```json
|
| 51 |
+
{
|
| 52 |
+
"system": "You are a helpful AI assistant with access to the following tools: ...",
|
| 53 |
+
"prompt": "What's the current price of NVDA?",
|
| 54 |
+
"chosen": "<tool_call>{...}</tool_call>\n<tool_result>{...}</tool_result>\n**NVDA** is trading at ...",
|
| 55 |
+
"rejected": "NVDA stock is around $875 but I don't have real-time data...",
|
| 56 |
+
"metadata": {"category": "stock_price", "ticker": "NVDA"},
|
| 57 |
+
"id": "abc123def456"
|
| 58 |
+
}
|
| 59 |
+
```
|
| 60 |
+
|
| 61 |
+
## Use Case
|
| 62 |
+
|
| 63 |
+
Fine-tune models to:
|
| 64 |
+
1. **Recognize** when a tool is needed vs. when to answer from knowledge
|
| 65 |
+
2. **Call tools correctly** with proper parameters
|
| 66 |
+
3. **Interpret and present** tool results clearly
|
| 67 |
+
|
| 68 |
+
Compatible with DPO, IPO, KTO, and other preference optimization frameworks.
|
| 69 |
+
|
| 70 |
+
## License
|
| 71 |
+
|
| 72 |
+
Apache 2.0
|