| --- |
| license: cc-by-4.0 |
| task_categories: |
| - text-generation |
| - reinforcement-learning |
| language: |
| - en |
| - zh |
| - ko |
| - ja |
| - fr |
| - ar |
| - it |
| - de |
| pretty_name: Flash Coding - logitly.ai |
| size_categories: |
| - 1K<n<10K |
| --- |
| # Flash Coding |
|
|
|  |
|
|
| > [!CAUTION] |
| > Uploading these datasets to proprietary AI (e.g., asking a question that includes them) can get your account suspended: the AI vendor can misclassify the chain of thought in these records as reverse engineering or duplicating model outputs. |
|
|
| Provider: [logitly.ai](https://logitly.ai) - Agentic data you can trust |
|
|
| Flash Coding is a dataset of agentic coding traces: an AI coding agent working on tasks in real source repositories. |
|
|
| Every trace is produced under a controlled runtime: |
|
|
| - Sandboxed environment, pinned repository, fixed working directory |
| - A restricted toolset and operating mode: `work`, `plan`, or `cplan` |
|
|
| An automated verification protocol validates every record before it enters the dataset. |
|
|
| Read the full specification: [specification.pdf](specification.pdf). |
|
|
| The dataset has two parts that share a single record schema: |
|
|
| | Part | Content | Training method | |
| |---|---|---| |
| | [SFT](https://huggingface.co/datasets/logitly-ai/flash-coding-sft-v1) | One demonstration trace per task | Supervised fine-tuning | |
| | [DPO](https://huggingface.co/datasets/logitly-ai/flash-coding-dpo-v1) | A chosen and a rejected trace of the same task | Direct preference optimization | |
|
|
| ## Trace Format |
|
|
| A trace is a list of messages in the style of the Anthropic Messages API: each message has a `role` (`user` or `assistant`) and a `content`, a list of typed blocks. Messages strictly alternate between the two roles, starting with `user` and ending with `assistant`. |
|
|
| | Block | Content | |
| |---|---| |
| | `text` | The task in the first user message; the assistant's narration between tool calls, or the final summary. | |
| | `thinking` | The agent's private reasoning for the turn. Plain text, never encrypted or summarized. | |
| | `tool_use` | A request to run one tool from `toolset`; `input` is a JSON object. | |
| | `tool_result` | The outcome of one `tool_use`, matched by id; `is_error` marks a failed call. | |
|
|
| Fixed shapes: |
|
|
| - The first user message opens with one `text` block: the task. Every later user message carries only `tool_result` blocks, one per `tool_use` of the preceding assistant message. |
| - Each assistant message holds one `thinking` block, then any mix of `tool_use` and `text` blocks in generation order. The last message closes the trace with a `text` block: a summary written for the user. |
|
|
| What a trace does not contain: |
|
|
| - **System prompt.** Not recorded: the trace begins with the task itself. |
| - **Media.** The four block types above are the only kinds; there are no image or document blocks. |
| - **Not distillation.** No `signature` and no `redacted_thinking`: the traces are not distilled from Anthropic models. |
| - **Tool schemas.** The trace carries the calls, not the definitions: `toolset` names the tools. |
|
|
| ## Tools |
|
|
| | Tool | Description | |
| |---|---| |
| | `glob` | Match paths against a glob pattern. | |
| | `grep` | Search file contents using regex. | |
| | `read` | Read contents from a file. | |
| | `write` | Write contents to a file, creating or overwriting it entirely. | |
| | `edit` | Replace a specific string in a file with new content. | |
| | `bash` | Execute a shell command. | |
|
|
| ## Usage |
|
|
| - **SFT**: each record carries a single trace, used for supervised training / fine-tuning. |
| - **DPO**: each record carries a `chosen` and a `rejected` trace of the same task, used for preference training: the model learns to prefer `chosen` over `rejected`. |
|
|
| ## Characteristics |
|
|
| - **System prompt.** Not given; it can be restored from the context fields. |
| - **Thinking.** Every assistant message carries a `thinking` block, so the traces carry the complete chain of thought. At training time, a `thinking` block can be dropped when it is short, so the same traces also serve adaptive thinking. |
| - **Language.** Tasks are written in multiple languages. In a positive trace, the assistant replies in the task's language; the language of `thinking` is not defined. |
| - **Error flag.** `is_error=true` marks a call that did not exit 0, which can be the ordinary outcome of a call. Keeping these results is deliberate: the model learns to recover from failures over long-horizon work. |
| - **Plan mode.** In `plan` mode, writing files is strictly forbidden, and `/tmp` is not writable either; the one exception is `/dev/null`. The mode gates use, not existence: the toolset is the same in every mode. |
|
|
| ## Evaluation |
|
|
| Flash Coding is training data: it carries no scores and no task-success labels, so the effect of training on it cannot be read off the data itself. Measure the effect on external benchmarks; we recommend DeepSWE and Terminal-Bench. The dataset shares no tasks with either benchmark: zero overlap, zero contamination. |
|
|
| ## Contact |
|
|
| Need a dataset tailored to your tasks? Email research@logitly.ai |
