daVinci-Dev: Agent-native Mid-training for Software Engineering
Paper • 2601.18418 • Published • 126
messages list | tools string | source_dataset string | uuid string | instance_id string | trajectory_id string |
|---|---|---|---|---|---|
[
{
"role": "assistant",
"content": "# Repository Context\nName: henricm/ha-ferroamp\nDescription: Ferroamp MQTT Home Assistant sensors for EnergyHub, SSO, ESM and ESO\n\n# Pull Request\n## chore: cleanup duplication in tests\n\n# Relevant Files Found\n## tests/test_init.py\n```\nimport uuid\nfrom unittest.mo... | GAIR__daVinci-Dev-ctx | 41 | |||
[
{
"role": "assistant",
"content": "# Repository Context\nName: open-mmlab/mmpretrain\nDescription: OpenMMLab Pre-training Toolbox and Benchmark\n\n# Pull Request\n## [Feature] Add ResNeSt configs\nHi，\r\n\r\nThese configs give acc of 80.41 for ResNeSt50 while the [original Pytorch code](https://github.com/z... | GAIR__daVinci-Dev-ctx | 332 | |||
[
{
"role": "assistant",
"content": "# Repository Context\nName: reazon-research/ReazonSpeech\nDescription: Massive open Japanese speech corpus\n\n# Issue\n## TypeError in load_model: unexpected keyword argument 'resume_download'\n## 問題の説明\nHowToガイドの「NeMoモデルで日本語の音声を認識する」を試した際にエラーが発生しました。 \n`load_model`関数を呼び出... | GAIR__daVinci-Dev-ctx | 56 | |||
[{"role":"assistant","content":"# Repository Context\nName: geopython/pycsw\nDescription: pycsw is a(...TRUNCATED) | GAIR__daVinci-Dev-ctx | 100 | |||
[{"role":"assistant","content":"# Repository Context\nName: google/adk-python\nDescription: An open-(...TRUNCATED) | GAIR__daVinci-Dev-ctx | 1183 | |||
[{"role":"assistant","content":"# Repository Context\nName: ilcardella/TradingBot\nDescription: Auto(...TRUNCATED) | GAIR__daVinci-Dev-ctx | 247 | |||
[{"role":"assistant","content":"# Repository Context\nName: getsentry/sentry-python\nDescription: Th(...TRUNCATED) | GAIR__daVinci-Dev-ctx | 2729 | |||
[{"role":"assistant","content":"# Repository Context\nName: kubernetes-sigs/inference-perf\nDescript(...TRUNCATED) | GAIR__daVinci-Dev-ctx | 83 | |||
[{"role":"assistant","content":"# Repository Context\nName: django/django\nDescription: The Web fram(...TRUNCATED) | GAIR__daVinci-Dev-ctx | 19293 | |||
[{"role":"assistant","content":"# Repository Context\nName: spotDL/spotify-downloader\nDescription: (...TRUNCATED) | GAIR__daVinci-Dev-ctx | 1287 |
End of preview. Expand in Data Studio
fan-shu/swe-davinci-ctx-midtrain
daVinci-Dev ctx-native (PR-derived) mid-training data, linearized for training
Qwen3-style code models. Rendered from GAIR/daVinci-Dev ctx-native/llm_enhanced_prs
following the daVinci Task-5 Markdown layout (Repository Context / Issue / Pull Request /
Relevant Files Found / Edits with search-replace blocks).
Configs
- cpt: continued-pretraining.
messages = [assistant: <full PR document>]; train on the whole document (loss_mask assistant=true). - sft: SFT.
messages = [user: <context>, assistant: <edits>]; loss on the edits only.
Source: https://huggingface.co/datasets/GAIR/daVinci-Dev (paper: https://arxiv.org/abs/2601.18418)
- Downloads last month
- 109