| --- |
| language: |
| - en |
| license: cc-by-4.0 |
| size_categories: |
| - 1K<n<10K |
| task_categories: |
| - text-generation |
| tags: |
| - synthetic |
| - reasoning |
| - agentic |
| - tool-use |
| - multi-turn |
| - sft |
| - grpo |
| pretty_name: Agentic Synth Reasoning |
| --- |
| |
| # Agentic Synth Reasoning |
|
|
| Synthetic reasoning traces for multi-step agentic behavior, generated incrementally from [nvidia/Nemotron-SFT-Agentic-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Agentic-v2). |
|
|
| ## Dataset Summary |
|
|
| Each record contains a multi-turn agentic conversation with **synthetically generated reasoning traces** (`reasoning_content`) attached to each assistant turn. The reasoning was produced **incrementally** — for each assistant message, only the conversation prefix up to that point was fed to the reasoning model, ensuring **zero future leakage**. |
|
|
| - **3,615 records** across **457 unique domains** |
| - **17,882 reasoning turns** (99.9% structured format) |
| - Average **3,002 chars** per reasoning trace |
| - Tool definitions from the source dataset included in reasoning context |
|
|
| ## Generation Method |
|
|
| 1. **Source**: Streamed conversations from `interactive_agent`, `tool_calling`, and `search` splits of Nemotron-SFT-Agentic-v2 |
| 2. **Incremental reasoning**: For each assistant message at position `i`, only `messages[0..i-1]` + tool definitions were fed to the model — making future leakage physically impossible |
| 3. **Structured format**: Reasoning follows a `### Situation Assessment` → `### Decision Rationale` structure with `●` (confirmed), `◐` (inferred), `○` (uncertain) confidence markers |
| 4. **Temperature escalation retry**: Empty/None responses retried at increasing temperatures (0.7 → 0.85 → 1.0) |
| 5. **Models used**: GLM-5.2 (via Ollama), DeepSeek V4 Flash (via OpenRouter) |
|
|
| ## Data Format |
|
|
| ```python |
| { |
| "messages": [ |
| {"role": "system", "content": "..."}, |
| {"role": "user", "content": "..."}, |
| {"role": "assistant", "content": "...", "reasoning_content": "### Situation Assessment\n● ..."}, |
| {"role": "tool", "content": "...", "tool_call_id": "..."}, |
| ... |
| ], |
| "domain": "theme park attraction management", |
| "source_dataset": "nvidia/Nemotron-SFT-Agentic-v2", |
| "original_model": "deepseek/DeepSeek-V3.2", |
| "num_original_messages": 12, |
| "num_reasoning_turns": 5, |
| "generated_at": "2026-07-04T16:55:49.744059+00:00" |
| } |
| ``` |
|
|
| ## Validation |
|
|
| | Check | Result | |
| |---|---| |
| | Format compliance | 99.9% (17,864/17,882 turns) | |
| | Future leakage | 0 confirmed (incremental generation) | |
| | Corrupted fields | 0 | |
| | Duplicates | 0 | |
|
|
| ## Intended Use |
|
|
| Designed as an SFT warmup dataset for GRPO-based LLM evaluator training. The reasoning traces teach models to: |
| - Decompose user requests in agentic contexts |
| - Select appropriate tools from available options |
| - Justify decisions with policy/constraint awareness |
| - Assess risks and edge cases before acting |
|
|
| ## Source Data |
|
|
| Derived from [nvidia/Nemotron-SFT-Agentic-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Agentic-v2). Original conversations preserved; only `reasoning_content` fields are synthetically generated. |
