---
license: apache-2.0
language:
- en
- zh
task_categories:
- text-generation
- question-answering
tags:
- reasoning
- chain-of-thought
- sft
- distillation
- finance
- multi-agent
size_categories:
- 100K<n<1M
---

# tejeshbhalladhanyog/sa-finance-reasoning-mix

A merged reasoning-distillation dataset combining a custom multi-agent
financial-reasoning pipeline with a stratified slice of GLM-5.1's
general-domain reasoning data.

## Composition

| source / subset | rows |
|---|---|
| `sa_pipeline_qwen_max` (decomposition + cluster_generation + mapreduce_single) | 153,923 |
| `main` | 80,000 |
| `PHD-Science` | 20,000 |
| `Multilingual-STEM` | 20,000 |
| `Math` | 15,000 |
| **total** | **288,923** |

## Format

Each row is a single conversation in standard messages format:

```json
{
  "messages": [
    {"role": "system",    "content": "..."},
    {"role": "user",      "content": "..."},
    {"role": "assistant", "content": "<think>\n...reasoning trace...\n</think>\n\n...final answer..."}
  ],
  "source": "sa_pipeline_qwen_max",
  "source_subset": "cluster_generation",
  "domain": "compliance",
  "stage": "cluster_generation",
  "question_id": "compliance.trade_surveillance_market_abuse.00124.635cf2.iter1",
  "teacher_model": "qwen3.6-max-preview",
  "input_tokens": 12345,
  "output_tokens": 6789
}
```

The assistant turn always carries reasoning as a leading
`<think>...</think>` block followed by the final answer, suitable for
SFT on chat templates that natively support reasoning blocks.

## Sources

- **SA Pipeline (qwen3.6-max-preview).** Multi-agent finance reasoning
  traces produced by an internal pipeline (decomposer → 3-7 cluster
  generators → map-reduce synthesizer). Domains span accounting, banking,
  trading, compliance, derivatives, structured credit, insurance, PE/VC,
  research, IR, tax, infra, real estate, etc.
- **GLM-5.1-Reasoning-1M-Cleaned.** Reformatted slice of
  `Jackrong/GLM-5.1-Reasoning-1M-Cleaned` covering `main`, `PHD-Science`,
  `Multilingual-STEM`, and `Math` subsets — sampled to retain general
  reasoning ability across domains.

## Loading

```python
from datasets import load_dataset
ds = load_dataset("tejeshbhalladhanyog/sa-finance-reasoning-mix", split="train", streaming=True)
for ex in ds:
    print(ex["messages"])
    break
```

## License

`apache-2.0` (inherited from the GLM-5.1-Reasoning-1M-Cleaned source for
the GLM portion; SA pipeline portion is released under the same terms).