lihaoxin2020/agentic-search-refiner-shortform-curated-v1

---
license: other
task_categories:
- question-answering
language:
- en
pretty_name: Agentic Search Refiner Shortform Curated v1
size_categories:
- 1K<n<10K
---

# Agentic Search Refiner Shortform Curated v1

Curated short-form QA mixture for `/home/nvidia/workspace/dr-tulu/rl/open-instruct/run_train_refiner_shortform_api_agent.sh`.

This dataset is normalized to the ASearcher-compatible schema expected by `rlvr_tokenize_asearcher_v1`:

- `question`: user query
- `answer`: exact reference answer as a string
- `source`, `subset`, `source_id`, `metadata`: provenance/debug fields

## Composition

```json
{
"asearcher_lrm_multihop": 300,
"asearcher_lrm_webwalker_chain": 300,
"webaggregatorqa_train": 300,
"webexplorerqa": 100,
"webshaper": 500
}
```

Total rows: `1500`.

## Curation

- WebShaper: all rows with non-empty question/answer retained.
- WebExplorerQA: all rows with non-empty question/answer retained.
- ASearcher LRM web/search-hard slice: 150 English-like `train_v1:webwalker_hard` and 150 English-like `chain_qa_fuzzy` rows, with approximate deduplication.
- ASearcher LRM multi-hop slice: 200 unlabeled LRM multi-hop-style rows and 100 `train_v1:compose_group_qa` rows, with approximate deduplication.
- WebAggregatorQA train: 300 rows with scalar answers, real URL context, no placeholder/sample-only answers, and topic/answer repetition caps.

## Intended Use

Use as a single JSONL/Hugging Face dataset in the short-form API-agent refiner RL run:

```bash
DATASET_LIST="lihaoxin2020/agentic-search-refiner-shortform-curated-v1 1.0" DATASET_SPLIT=train \
bash /home/nvidia/workspace/dr-tulu/rl/open-instruct/run_train_refiner_shortform_api_agent.sh
```

Local mirror:

```bash
DATASET_LIST="/home/nvidia/workspace/data/agentic_search_refiner_shortform_curated_v1.jsonl 1.0" DATASET_SPLIT=train \
bash /home/nvidia/workspace/dr-tulu/rl/open-instruct/run_train_refiner_shortform_api_agent.sh
```

See `manifest.json` for exact counts, filter notes, and representative examples.