Zeta Dataset (Python Generator)

This repository provides a Python + uv workflow for generating an edit‑prediction dataset. You supply short code excerpts ("seeds") and generate training examples with OpenAI GPT‑5.2. The resulting dataset keeps Zeta‑style markers and is suitable for fine‑tuning models such as qwen3-coder-30b-a3b-instruct.

Dataset Format

Each example is a Markdown file containing these tags, in this order:

• <events>: A short description of the user edit, including a unified diff.
• <input>: The original excerpt with editable region markers and one cursor marker.
• <output>: The edited excerpt with the same editable region markers and no cursor marker.
• <assertions> (optional): One short sentence validating the change.
• <labels>: Two comma-separated labels (edit location + intent).

Marker rules:

• <input> must contain exactly one <|editable_region_start|>, one <|editable_region_end|>, and one <|user_cursor_is_here|>.
• <output> must contain exactly one <|editable_region_start|> and one <|editable_region_end|>, and no <|user_cursor_is_here|>.
• Content outside the editable region must be identical between <input> and <output>.

Label format:

• Location: no-op | local-edit | non-local-edit
• Intent: add-imports | complete-implementation | complete-pattern | infer-intent | infer-refactor | unknown

Workflow

0. Install dependencies (uv)

uv sync

1. Extract seeds automatically

uv run python script/extract_seeds.py \\
  --api-repo /path/to/yard-api \\
  --app-repo /path/to/yard-app \\
  --out seeds \\
  --api-count 100 \\
  --app-count 100 \\
  --min-lines 8 \\
  --max-lines 16 \\
  --clean

If you want to clone over SSH instead of providing local paths:

uv run python script/extract_seeds.py \\
  --api-repo git@github.com:Hyperduality/yard-api.git \\
  --app-repo git@github.com:Hyperduality/yard-app.git \\
  --out seeds \\
  --api-count 100 \\
  --app-count 100 \\
  --min-lines 8 \\
  --max-lines 16 \\
  --clean \\
  --clone

When --clone is used, repos are cloned into repos/ (ignored by git).

This samples a variety of files from both repos (controllers, services, schemas, components, hooks, shared, etc.) into seeds/.

2. Generate examples with OpenAI

Set your API key (from .env or environment):

OPENAI_API_KEY=your_key_here

Optional:

• OPENAI_ORGANIZATION_ID
• OPENAI_PROJECT_ID

Run:

uv run python script/gen_examples.py --seeds seeds --out . --model gpt-5.2

This writes examples into train/ and eval/. Each model response can include multiple examples separated by a line containing only <<<EXAMPLE>>>.

3. Validate formatting

uv run python script/check_format.py

4. Generate JSONL datasets

uv run python script/gen_dataset.py

Outputs:

• train.jsonl
• eval.jsonl

Manual Seeds (Optional)

If you want to add manual seeds, use this format:

path: relative/path/to/file.ts
---
<raw excerpt here>

You can mix manual seed files with auto‑generated ones under seeds/.

Customizing Prompts

Edit script/prompt-template.md to change how examples are generated. Placeholders:

• {{output_path}}
• {{path}}
• {{excerpt}}

Notes

• Prompts, seeds, and generated datasets are ignored by default via .gitignore.
• This repo no longer includes Snowflake, DPO, or labeling pipelines.