SuperAGI/superscout-sft-search

SuperScout search corpus (SFT)

The supervised fine-tuning corpus behind SuperScout-7B, a 7B searcher that explores a repository, localizes the fault, writes a failing reproduction, and emits a structured handoff. The dataset contains 19,911 examples as built and frozen; six rows carrying malformed tool-call wrappers are dropped at load time, giving the 19,905 examples actually trained on. A further 9,478 examples (the third-best trace per issue) were held back as a shelf and never trained on.

The trained model is at SuperAGI/SuperScout-7B; the paper, code, and evaluation receipts are at TransformerOptimus/superscout.

What one example is

Each example is one complete search episode sliced from an agent trajectory: the agent explores a repository, localizes the fault, and writes a failing reproduction, cut at the point the search is complete, with a synthesized handoff-emission turn appended as the final supervised target. Within that synthesized turn:

• the file list is extracted deterministically from the trace;
• the reproduction record is copied verbatim;
• only the free-form notes are model-written, by the open-weights gpt-oss-120b.

Loss is taken on assistant turns only, so environment output never contributes a gradient.

Sources and filtering

Traces are sliced from three public, openly licensed (CC-BY-4.0) trajectory sets:

Source	Role
Open-SWE-Traces	Go / TypeScript / JavaScript + Python
SWE-rebench-openhands	Python (OpenHands scaffold trajectories)
SWE-Hero	Python (OpenHands scaffold trajectories)

Trajectories are success-filtered against the gold patch, keeping only traces that actually found the right files, then deduplicated to the two highest quality-ranked traces per issue so the corpus counts distinct bugs rather than retellings of the same fix. 97.7% of examples carry a verified reproduction.

Composition

Language	Examples	Share
Python	7,417	37.3%
Go	7,304	36.7%
TypeScript	3,932	19.7%
JavaScript	1,258	6.3%

Percentages are of the 19,911-example built set; the 6 load-time drops are not attributed to a language. One epoch is 382.4M tokens, packed into 12,723 blocks of 32k tokens at 91.7% fill.

Contamination control

A 23-repository blocklist, decided before any training data was built, excludes every repository appearing in SWE-bench Pro (11 repositories) or SWE-bench Verified (12 repositories) from all training and calibration data. In addition, all 450 issues of a held-out evaluation vault are excluded. Both exclusions were verified by a programmatic gate on the frozen file (zero hits), not assumed. The blocklist ships with the paper repository as data/blocklist.json.

Citation

@misc{superscout2026,
  title         = {Scrouting: Cost-Aware Routing of Coding Agents by Scouting the Repository First},
  author        = {Ishaan Bhola and Adithyan Krishnan and Mukunda NS},
  year          = {2026},
  eprint        = {2608.04804},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SE}
}