---
configs:
- config_name: sft_balanced
  data_files:
  - split: train
    path: train-*.parquet
  - split: validation
    path: validation-*.parquet
task_categories:
- text-generation
language:
- en
tags:
- distillation
- sft
- reasoning
- tool-use
- multi-turn
- multi-teacher
- chain-of-thought
- thinking
- devin
- glm-5.2
- agent
size_categories:
- 1K<n<10K
license: other
---

# Devin CLI Reasoning Distillation Dataset

A distillation dataset built from **Devin CLI** session traces, containing the model's internal
**reasoning traces** (chain-of-thought / thinking), user prompts, assistant answers, and tool calls.
The dataset is formatted to be directly compatible with SFT training pipelines that expect
OpenAI-style message lists with a `reasoning_content` field.

## Dataset Summary

| | |
|---|---|
| **Total rows** | 2,632 (2,507 train / 125 validation) |
| **Rows with reasoning** | 2,507 / 2,507 (100% of train) |
| **Format** | Parquet (32 columns, matches `r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation`) |
| **Source** | Devin CLI `sessions.db` (local agent traces) |
| **Schema version** | 1.0.0 |

## Teacher Models

This is a **multi-teacher** distillation dataset. The dominant teacher is **GLM-5.2** (70%),
with contributions from Devin's own SWE-1.7 and several other frontier models:

| Model | Rows | % |
|-------|------|---|
| **glm-5-2** | **1,758** | **70.2%** |
| swe-1-7 | 405 | 16.2% |
| glm-5-3-flash-max | 168 | 6.7% |
| gpt-5-6-luna-xhigh | 69 | 2.8% |
| deepseek-v4-flash-max | 25 | 1.0% |
| gpt-5-6-luna-high | 24 | 1.0% |
| gemini-3-7-flash-high | 23 | 0.9% |
| glm-5-3-flash-low | 18 | 0.7% |
| claude-opus-5-high | 10 | 0.4% |
| claude-sonnet-5-high | 7 | 0.3% |

## Dataset Structure

Each row represents a conversation segment (anchored to a single user prompt) containing:

- The full message history (system prompts, user input, assistant responses, tool calls, tool results)
- The assistant's **reasoning trace** in the `reasoning_content` field — the model's internal
  chain-of-thought before producing its answer
- Tool calls in OpenAI function-calling format

### Key Columns

| Column | Type | Description |
|--------|------|-------------|
| `messages_json` | string | JSON array of messages, each with `content`, `reasoning_content`, `role`, `tool_calls`, `tool_call_id`, `name`, `trainable` |
| `teacher_model` | string | The model that generated the response |
| `teacher_provider` | string | Always `cognition` |
| `trace_kind` | string | `single_turn` or `multi_turn` |
| `n_assistant_turns` | int32 | Number of assistant turns in the segment |
| `n_tool_calls` | int32 | Number of tool calls in the segment |
| `domain` | string | Inferred domain (coding, reasoning, math) |
| `source_item_id` | string | `{session_id}#{segment_index}` for traceability |

### Message Format

Each message in `messages_json` follows this structure:

```json
{
  "content": "The visible answer or prompt text",
  "reasoning_content": "The model's internal thinking trace (chain-of-thought)",
  "role": "system|user|assistant|tool",
  "tool_calls": [{"id": "...", "type": "function", "function": {"name": "...", "arguments": "..."}}],
  "tool_call_id": "",
  "name": "",
  "trainable": true
}
```

- `trainable: true` for assistant messages (the target for distillation)
- `trainable: false` for system, user, and tool messages (context only)
- `reasoning_content` is populated for assistant messages that include a thinking trace

## Use Cases

- **SFT distillation**: Train a student model to reproduce the reasoning + answer patterns
- **Reasoning training**: Teach chain-of-thought reasoning with tool use
- **Agent training**: Multi-turn agentic workflows with tool calls
- **Multi-teacher distillation**: Blend knowledge from GLM-5.2, SWE-1.7, and other models

## Statistics

| Metric | Value |
|--------|-------|
| Sessions | 530 |
| Avg assistant turns per segment | 9.0 |
| Avg tool calls per segment | 9.7 |
| Single-turn segments | 2,479 (99%) |
| Multi-turn segments | 28 (1%) |

## Provenance

- **Source**: Devin CLI agent sessions running on a local Windows workstation
- **Collection period**: July 2026 – September 2026
- **Tasks**: Software engineering, code debugging, ML training, research, general Q&A
- **Extraction**: Conversations reconstructed from `sessions.db` message chain, split at user-turn
  boundaries, filtered to segments containing at least one reasoning trace

## License

This dataset contains traces from proprietary AI models accessed via the Devin CLI platform.
Usage is subject to the terms of service of the respective model providers.