Salvor-Hardin/Search-Distill-5K-ReAct

---
language:
- en
license: apache-2.0
size_categories:
- 1K-10K
task_categories:
- text-generation
- question-answering
tags:
- react
- search-distillation
- reasoning
- rag
- slm
configs:
- config_name: default
data_files:
- split: train
path: train.parquet

---

# TraceMind-Search-5K

**TraceMind-Search-5K** is a high-quality dataset featuring 5,000+ samples of search-augmented reasoning traces. It is specifically designed to train Small Language Models (SLMs) in the **ReAct (Reasoning + Acting)** pattern, focusing on precise information retrieval, noise filtering, and grounded synthesis.

## Dataset Highlights

- **Total Samples**: 4,997 entries in Parquet format.
- **File**: `train.parquet` (2.7 MB).
- **Pattern**: Strict adherence to the ReAct framework:
`User` -> `Thought (pre-search)` -> `<search>` -> `<|results|>` -> `Thought (post-search)` -> `Answer`.
- **Reasoning Depth**: Each entry includes an internal monologue that justifies the search query and critiques the retrieved results.
- **Synthesized Facts**: Contains up-to-date reasoning on synthetic "future" events (2025-2026) to test grounding performance.

## Methodology: Search Distillation

The dataset was generated using a "Search Distillation" methodology:
1. **Thought (Pre-search)**: The model identifies knowledge gaps and formulates a search strategy.
2. **Search**: Concise, keyword-based queries are generated.
3. **Synthesis**: The model reviews multiple sources, explicitly dismisses irrelevant "noise," and synthesizes a final answer.
4. **Grounding**: The final answer is strictly grounded in the provided `<|results|>` block.

## Use Cases

- **RAG Alignment**: Fine-tune models to better utilize external context and filter out search noise.
- **Tool-Use Training**: Train models to generate effective search queries.
- **SLM Reasoning**: Improve the "chain-of-thought" capabilities of smaller models through high-quality distillation traces.

## Dataset Structure

The data is provided in a single `.jsonl` file with a `text` field containing the full trace:

```json
{"text": "User: ... \nAssistant: Thought: ... \n<search>...</search> \n<|results|> ... <|/results|> \nAssistant: Thought: ... \nAnswer: ..."}
```

## Dataset Generation
This dataset consists of synthetic ReAct (Reasoning + Acting) traces generated using the open-weights **`moonshotai/kimi-k2-instruct-0905`** model.

The model was specifically prompted to produce high-quality noise-filtered trajectories for training Small Language Models (SLMs).

## License
This dataset and its curated structure are released under the **Apache License 2.0**. The source data is derived from Kimi K2, which operates under Moonshot AI's Modified MIT License, allowing for broad open-source and commercial downstream use.