SERA Training Data
SERA training data consists of over 200,000 synthetic coding agent trajectories generated using the Soft Verified Generation (SVG) method. Unlike previous approaches that require test suites or bug injection pipelines, SVG generates training data from any codebase by using line-level patch matching for verification. The data covers 121 Python repositories and 51 bug types, providing diverse training signal for coding agents.
Dataset variants
Each dataset variant is split into T1 (first rollout) and T2 (second rollout) trajectories:
| Dataset | Description | Size | Teacher Model |
|---|---|---|---|
| Sera-4.5A-Lite | Compact version for faster training | ~72k trajectories | GLM-4.5-Air |
| Sera-4.5A-Full | Complete dataset | ~140k trajectories | GLM-4.5-Air |
| Sera-4.6-Lite | Generated with stronger teacher | ~72k trajectories | GLM-4.6 |
- T1 trajectories: First rollout where the teacher makes changes starting from a randomly selected function
- T2 trajectories: Second rollout where the teacher reproduces the change given only the synthetic PR description
Data generation method
The Soft Verified Generation (SVG) method generates training data through two rollouts:
- First rollout: A teacher model is prompted to make a change to a codebase starting from a randomly selected function, producing a trajectory and patch
- Synthetic PR generation: The trajectory is converted into a synthetic pull request description
- Second rollout: The teacher attempts to reproduce the original change given only the PR description
- Soft verification: The two patches are compared using line-level recall
Trajectories are selected based on verification threshold (recall score), with options ranging from completely unverified (r=0) to hard-verified (r=1).
Data format
Each trajectory contains:
- Task description: The synthetic PR or bug prompt that initiated the rollout
- Agent actions: A sequence of tool calls (view, edit, bash) and observations
- Final patch: The code diff produced by the agent
The format matches the SWE-agent tool format. See the SERA training repository for detailed schema documentation.
Access
All datasets are available on HuggingFace under the ODC-BY license in the Open Coding Agents collection.
Generating your own data
The SERA training repository contains the full pipeline for generating training data from your own codebases. This enables specialization to private repositories without exposing code to third parties.
See the SERA training repository for instructions on data generation and fine-tuning.