AmanPriyanshu/reasoning-sft-JSON-structuring-and-correcting

---
license: apache-2.0
task_categories:
- text-generation
- question-answering
language:
- en
- ja
tags:
- reasoning
- sft
- chain-of-thought
- tool-calling
- structured-output
- json
- json-repair
- agent
size_categories:
- 100K<n<1M
---

# JSON Structuring and Correcting (Reasoning SFT)

Combined dataset of 508K rows for training LLMs on structured output tasks with reasoning traces, sourced from two datasets:

## Sources

### tool_calling.parquet (488,461 rows)
Converted from [vericava/sft-tool-calling-structured-output-v1](https://huggingface.co/datasets/vericava/sft-tool-calling-structured-output-v1). Multi-turn tool calling and structured output tasks including tool invocations, tool results, and final assistant responses. Includes English and Japanese content.

### json_repair.parquet (19,645 rows)
Converted from [kth8/json-repair](https://huggingface.co/datasets/kth8/json-repair). JSON formatting and repair tasks where the model corrects malformed JSON into valid, properly indented output.

## Format

Each row has three columns:

- **`input`** — list of dicts with role/content conversation turns (system prompt includes tools/schema where applicable)
- **`response`** — response string with `<think>` reasoning block followed by the structured output
- **`source`** — `N/A`

## Conversion

- Tool-calling: system prompt combines model identity + available tools + output schema; tool calls and results preserved as context turns; 1,322 rows dropped (empty targets)
- JSON repair: system + user turns as input, assistant response as output; 0 rows dropped
- Both: short task-focused reasoning injected in think blocks; validated exactly 1 open and 1 close think tag per response

## Usage

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

repo = "AmanPriyanshu/reasoning-sft-JSON-structuring-and-correcting"

for name in ["tool_calling.parquet", "json_repair.parquet"]:
path = hf_hub_download(repo_id=repo, filename=name, repo_type="dataset")
table = pq.read_table(path)
print(f"{'='*80}")
print(f"{name}: {len(table)} rows")
print(f"{'='*80}")
i = random.randint(0, len(table) - 1)
row = {col: table.column(col)[i].as_py() for col in table.schema.names}
print(f"\n[source] {row['source']}")
print(f"\n[input] ({len(row['input'])} turns)")
for t in row["input"]:
preview = t["content"][:250] + ("..." if len(t["content"]) > 250 else "")
print(f" {t['role']}: {preview}")
rp = row["response"][:800]
if len(row["response"]) > 800:
rp += "..."
print(f"\n[response]\n{rp}\n")
```

## License

Apache 2.0 — inherited from both source datasets.