AmanPriyanshu/tool-reasoning-sft-TOOLS-hermes-reasoning-tool-style-data-cleaned-rectified-115k

---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- deep-research
- reasoning
- tool-calling
- agentic
- multi-hop
- search
size_categories:
- 1K<n<10K
---

---

## Agentic Tool-Use SFT Mix

111,295 additional multi-turn agentic trajectories across four task families, following the same strict reasoning + tool-call FSM format. Combined with the original 3,827 deep-research trajectories, the dataset totals **115,122 samples**.

### Distribution

| Category | Samples | Full | Compact |
|---|---|---|---|
| Deep Research (original) | 3,827 | 100% | — |
| Multi-Turn Tool Orchestration | 45,776 | 54% | 46% |
| Deep Research | 34,282 | 71% | 29% |
| Codebase Retrieval | 17,473 | 69% | 31% |
| Database Interaction | 13,764 | 69% | 31% |
| **Total** | **115,122** | | |

### Schema

Two columns: `messages` (JSON string — list of role/content dicts) and `source` (category label).

### Cleaning

All trajectories validated against the strict FSM. Stray turns stripped, missing reasoning bridges inserted, consecutive reasoning merged. ~11k trajectories required at least one repair.

```
system → user → reasoning → tool_call → tool_output → reasoning → tool_call → ... → reasoning → answer
```

## Validated Transitions

```
system → user
user → reasoning
reasoning → tool_call | answer
tool_call → tool_output
tool_output → reasoning
answer → user (multi-turn only)
```

## Usage

```py
import json, random
from huggingface_hub import hf_hub_download
import pyarrow.parquet as pq

REPO = "AmanPriyanshu/tool-reasoning-sft-hermes-reasoning-tool-style-data-cleaned-rectified-115k"

FILES = ["compiled_data.parquet", "data.parquet"]

for fname in FILES:
print("=" * 70)
print(f"Downloading {fname}...")
local = hf_hub_download(REPO, fname, repo_type="dataset")

t = pq.read_table(local)
print(f"Rows: {t.num_rows:,} | Columns: {t.column_names}")

idx = random.randint(0, t.num_rows - 1)
row = {col: t.column(col)[idx].as_py() for col in t.column_names}

msgs = json.loads(row["messages"])
meta = {k: v for k, v in row.items() if k != "messages"}
print(f"\nRow {idx} | meta={meta} | {len(msgs)} turns")
print(f"Roles: {' -> '.join(m['role'] for m in msgs[:20])}{'...' if len(msgs) > 20 else ''}\n")

for m in msgs:
content = m["content"]
if m["role"] == "system":
content = content[:200] + "..."
elif len(content) > 300:
content = content[:300] + "..."
print(f"[{m['role']}]\n{content}\n")

print()
```

## License

Apache-2.0