0xclaw-ai/budgetrouter-hackrep

BudgetRouter HackRep Phase Controls

This is a redacted release of the data constructed for BudgetRouter from HackRep. It contains 56,311 selected projects and 394,177 phase examples. Each project contributes seven examples: research, idea, selection, planning, coding, testing, and documentation.

The targets are heuristic weak-supervision labels derived from project metadata and README features. They are not measured optimal budgets or human annotations.

• Dataset preparation and training
• BudgetRouter runtime
• Trained adapter
• DeadlineFramer
• 0xClaw
• Paper / preprint

Source, license, and attribution

HackRep was created by Sjoerd Halmans, Lavinia Paganini, Alexander Serebrenik, and Alexander Nolte. The official Zenodo dataset record identifies its license as Creative Commons Attribution 4.0 International (CC BY 4.0). This is the dataset license; the accompanying research paper has a separate license.

We distribute the collection and our original annotations and documentation under CC BY 4.0, with the scope described in NOTICE.md. Retain the attribution, license link, and indication of modifications when redistributing. See LICENSE and dataset-provenance.json.

Source-project README excerpts and other third-party material may have separate rights. The collection license does not relicense those projects. The preserved license metadata reports 47,029 projects without a license value and 405 with NOASSERTION; absence of a value does not establish permission or prohibition. The other 8,877 projects have a reported SPDX identifier, which is not an independent verification of the license's scope. Use project_provenance.jsonl to locate the source and its recorded license. This release does not certify that every source project is available for unrestricted reuse.

The raw upstream parquet file was checked against the official Zenodo MD5 checksum. The original parquet, unredacted generated files, backups, model weights, and local training logs are not included.

Construction and splits

The source contains 100,356 projects. The existing preparation pipeline filters projects using README length, commit count, duration, repository status, and language requirements; it retains 56,311 projects. It extracts features, assigns heuristic phase controls, and splits by project with seed 20260515. All phases of a project remain in one split, with no project overlap.

Split	Projects	Phase examples
Train	50,669	354,683
Validation	2,819	19,733
Test	2,823	19,761
Total	56,311	394,177

The flat configuration contains the same phase examples as sft in another representation. Do not count both configurations as independent samples. The projects and provenance configurations each expose an inventory split covering all projects; their split field records the actual project partition.

Files and schema

Files	Contents
phase_budget_sft_*.jsonl	messages with system/user/assistant roles, plus metadata containing project ID, phase, split, labels, and weak-supervision scores
phase_budget_flat_*.jsonl	Project ID, phase, split, input text, target control string, label details, and numeric/categorical features
clean_projects.jsonl	Selected project metadata, redacted description and README excerpt, extracted features, split, and phase labels
project_provenance.jsonl	Project ID, source URL, split, recorded upstream license fields, and source DOI
manifest.json	Original construction settings and counts, with local paths redacted
sanitization-report.json	Redaction counts and original/export checksums; no original text or credential values
dataset-provenance.json	Source attribution, license evidence, transformation summary, and checkpoint sampling configuration
release-files.json	Exact publication file list, sizes, and SHA-256 checksums

Example target:

<PHASE_CODING><TOKEN_1.0X><TIME_HIGH><THINK_LOW><REPAIR_1>

Source excerpts can contain multiple languages. The instructional framing, control labels, and release documentation are in English.

Load with Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("0xclaw-ai/budgetrouter-hackrep", "sft")
train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]

Use streaming=True to iterate without materializing the dataset in an Arrow cache. Other configuration names are flat, projects, and provenance.

Use with the existing trainer

hf download 0xclaw-ai/budgetrouter-hackrep --repo-type dataset --local-dir ./data/budgetrouter-hackrep --include phase_budget_sft_train.jsonl --include phase_budget_sft_val.jsonl --include phase_budget_sft_test.jsonl

In your local copy of the training YAML, set:

data:
  train_file: ./data/budgetrouter-hackrep/phase_budget_sft_train.jsonl
  val_file: ./data/budgetrouter-hackrep/phase_budget_sft_val.jsonl
  max_seq_len: 1024
  max_train_samples: 50000
  max_val_samples: 4096

Keep the other training settings from the training repository. The existing loader creates fresh byte-offset indices for the downloaded files.

The released 50k adapter was trained before this redaction pass. Its recorded configuration samples at most 50,000 training examples and 4,096 validation examples using seed 20260515. These are example counts, not project counts. The full 394,177-example corpus is larger than the checkpoint's sampled inputs. Publishing this redacted copy does not retrain the adapter, and its weights remain unchanged. Training on redacted text is not a claim of bitwise reproduction of the earlier checkpoint.

Redaction and limitations

The export replaces recognized API-key formats, secret assignments, bearer tokens, private-key material, credential-bearing URLs, and local machine paths. It changes input text in 5,208 SFT examples and preserves sample order, project splits, metadata, numeric features, labels, and assistant control strings. Corresponding flat examples and project records are redacted as well.

The seven original redacted JSONL exports are packaged byte-for-byte. The provenance sidecar is additional metadata and does not alter training inputs. Stale byte-offset indices are excluded. Automated scanning found no remaining candidates within the checked patterns; it cannot establish that every possible secret or personal detail has been removed. Public project identifiers and source URLs remain for attribution and reproducibility. Source text should be treated as untrusted input, not as instructions to execute.

These labels inherit the preparation heuristics and source-selection biases. They do not guarantee a task's runtime, token use, or completion. Do not use the test split to tune the routing policy and then report it as an independent evaluation.

Citation

@dataset{halmans2025hackrep,
  author = {Halmans, Sjoerd and Paganini, Lavinia and Serebrenik, Alexander and Nolte, Alexander},
  title = {HackRep: A Large-Scale Dataset of GitHub Hackathon Projects},
  year = {2025},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.17572684},
  url = {https://doi.org/10.5281/zenodo.17572684}
}

Please also cite the linked 0xClaw paper when using the derived phase-control annotations. Report the dataset revision and the redaction status in experiments.