README
SWE-Pruner Pro Training Corpus
Training data for SWE-Pruner Pro, a lightweight in-agent context pruner for long-horizon coding agents.
Overview
- 22,609 samples of line-level keep/prune annotations on coding agent tool responses
- Annotated by Claude Sonnet 4.6 with reasoning and confidence
- Source trajectories from ByteDance-Seed/Multi-SWE-bench_trajs
Fields
| Field | Description |
|---|---|
| instance_id | SWE-bench instance identifier |
| step_idx | Step index within the trajectory |
| tool_call | The agent tool call that produced the response |
| tool_response | Raw tool response (to be pruned) |
| kept_frags | Line ranges marked as "keep" |
| total_lines | Total lines in tool_response |
| history | Preceding conversation context |
| next_turn | The agent's next action (used to verify pruning doesn't lose info) |
| _labeler | Annotation tool (claude) |
| _model | Specific model used for annotation |
| _source | Source trajectory dataset |
| _skeleton | Whether this is a skeleton-format sample |
| _reasoning | Natural language explanation of keep/prune decisions |
| _confidence | Annotator confidence level |
Usage
from datasets import load_dataset
dataset = load_dataset("ayanami-kitasan/swe-pruner-pro-training-corpus")
For training and inference usage, please refer to the GitHub repository.
Links
- Paper: arXiv 2607.18213
- Code: GitHub
- Model (Qwen3-Coder-Next head): ayanami-kitasan/swe-pruner-pro-qwen3-coder-next-head
- Model (MiMo-V2-Flash head): ayanami-kitasan/swe-pruner-pro-mimo-v2-flash-head
Contributors
ayanami-kitasan
6 commits
Similar repos
Linked in READMEs
README
SWE-Pruner Pro Training Corpus
Training data for SWE-Pruner Pro, a lightweight in-agent context pruner for long-horizon coding agents.
Overview
- 22,609 samples of line-level keep/prune annotations on coding agent tool responses
- Annotated by Claude Sonnet 4.6 with reasoning and confidence
- Source trajectories from ByteDance-Seed/Multi-SWE-bench_trajs
Fields
| Field | Description |
|---|---|
| instance_id | SWE-bench instance identifier |
| step_idx | Step index within the trajectory |
| tool_call | The agent tool call that produced the response |
| tool_response | Raw tool response (to be pruned) |
| kept_frags | Line ranges marked as "keep" |
| total_lines | Total lines in tool_response |
| history | Preceding conversation context |
| next_turn | The agent's next action (used to verify pruning doesn't lose info) |
| _labeler | Annotation tool (claude) |
| _model | Specific model used for annotation |
| _source | Source trajectory dataset |
| _skeleton | Whether this is a skeleton-format sample |
| _reasoning | Natural language explanation of keep/prune decisions |
| _confidence | Annotator confidence level |
Usage
from datasets import load_dataset
dataset = load_dataset("ayanami-kitasan/swe-pruner-pro-training-corpus")
For training and inference usage, please refer to the GitHub repository.
Links
- Paper: arXiv 2607.18213
- Code: GitHub
- Model (Qwen3-Coder-Next head): ayanami-kitasan/swe-pruner-pro-qwen3-coder-next-head
- Model (MiMo-V2-Flash head): ayanami-kitasan/swe-pruner-pro-mimo-v2-flash-head
Similar repos
Linked in READMEs
Contributors
ayanami-kitasan
6 commits