datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
semantic-repair-routing
semantic-repair-routing
The supervised pairs that train
SemanticRepair-270M:
a message somebody actually wrote, and the requests inside it restated
plainly, one per line. 84,819 pairs in five languages, plus 2,515 in
Italian and English aimed at what the model used to refuse.
It teaches one narrow thing. An embedding router compares a question with
the description of every capability it can reach. People do not write the
way capabilities are described — they hedge, they… See the full description on the dataset page: https://huggingface.co/datasets/Gramscii-IT/semantic-repair-routing.python-program-repair-training-pool
Python program-repair training pool
A pool of public data for training a model to repair broken Python. Every row is a
program that does the wrong thing and the program that replaces it. It is a straight
collection of open datasets plus a rule-generated layer built from open functions, not a
new corpus: every row comes from one of the sources below, at the revision named, and
every row was put through an overlap filter against held-out material this pool is kept
separate from.… See the full description on the dataset page: https://huggingface.co/datasets/Emulated-Inc/python-program-repair-training-pool.openenv-python-repair
Python Repair Lab
An original OpenEnv curriculum of 1,200 deterministic Python function-repair episodes: 12 problem families, four distinct bug patterns per family, and 25 seeded case sets per pattern. There are 151,780 executable checks across the episodes. These are 48 repair patterns with data variants, not 1,200 unrelated algorithms. Tasks cover interval algorithms, rolling calculations, weighted statistics, stable deduplication, Unicode run-length encoding, Luhn checksums… See the full description on the dataset page: https://huggingface.co/datasets/Louistiti/openenv-python-repair.SWE-universe-repaired-bug-pilot-trajectories
SWE-universe repaired BugPilot trajectories
Combined trajectory artifacts for the Qwen3.6 + mini-swe-agent evaluation of VmaxRL/SWEUniverse-Repaired-Bugpilot.
This dataset contains one row per evaluated task in metadata.jsonl, plus per-task files under trajectories//. The combined set uses the main full eval and replaces the two original infra-failure rows with the clean infra rerun trajectories.
Summary:
rows: 804
effective attempts: 804
passes: 629
pass rate: 0.782338
infra… See the full description on the dataset page: https://huggingface.co/datasets/VmaxRL/SWE-universe-repaired-bug-pilot-trajectories.uncgpt-conversations-semantic-approved-1p25-repaired-paperclip
UncGPT — Semantic-Approved 1.25σ Conversations (Leak-Repaired)
The 1.25σ semantic-gate cohort with uncle-diary leakage repaired and normalized diary fields. The auditable replacement for the older paperclip_all_1803 source that an earlier audit flagged for visible diary leakage.
Part of the UncGPT NeurIPS 2026 Competition collection.
Config
approved_manifest (default): one row per approved conversation, with metadata + path back to the full-schema JSON.… See the full description on the dataset page: https://huggingface.co/datasets/Reza2kn/uncgpt-conversations-semantic-approved-1p25-repaired-paperclip.swerl-tmax-15k-repairs-gpt-5-6-sol
swerl-tmax-15k task repairs (gpt-5-6-sol)
Proposed repairs for defective tasks in hamishivi/swerl-tmax-15k, generated from
the audit labels in
wAI-org/swerl-tmax-15k-rubric-gpt-5-6-sol.
These repairs are UNVALIDATED
No repair here has been executed, and none has been shown to be solvable.
Every repair was checked mechanically — valid bash, still writes a reward,
does not delete a path the instruction needs. None was checked empirically.
A hardened verifier can be… See the full description on the dataset page: https://huggingface.co/datasets/wAI-org/swerl-tmax-15k-repairs-gpt-5-6-sol.Palace-Config-Repair
Palace Configuration Repair Benchmark
Execution-graded repair tasks for configuration files of
Palace, an open-source finite-element solver for
computational electromagnetics. Each task gives a model a perturbed Palace JSON configuration and
asks it to return a corrected one. A repair counts as correct only if it conforms to the schema,
is accepted by the solver, and reproduces the reference outputs of the original case when
Palace v0.14.0 runs it. A configuration that is valid… See the full description on the dataset page: https://huggingface.co/datasets/empirischtech/Palace-Config-Repair.SWEUniverse-Repaired-Indist-full-not-SWE-bench-pro-matched
VmaxRL/SWEUniverse-Repaired-Indist-full-not-SWE-bench-pro-matched
This dataset contains a 350-row subset selected from the Indist SWEUniverse training rows.
Selection policy: three-way repo overlap with Bugpilot and LM-Modify, deduped by repo plus introduction patch, then balanced round-robin across overlapping repos.
Rows: 350
Selected repos: 19
Deduped overlap capacity: 468
Source dataset: VmaxRL/SWEUniverse-Repaired-Indist-full-not-SWE-bench-pro-matched
repairllama-datasets
RepairLLaMA - Datasets
Contains the processed fine-tuning datasets for RepairLLaMA.
Instructions to explore the dataset
To load the dataset, you must define which revision (i.e., which input/output representation pair) you want to load.
from datasets import load_dataset
# Load ir1xor1
dataset = load_dataset("ASSERT-KTH/repairllama-datasets", "ir1xor1")
# Load irXxorY
dataset = load_dataset("ASSERT-KTH/repairllama-datasets", "irXxorY")
Citation
If you use… See the full description on the dataset page: https://huggingface.co/datasets/ASSERT-KTH/repairllama-datasets.TobiaSVG-repair
TobiaSVG Repair
TobiaSVG Repair contains 44,802 synthetic SVG repair pairs. Each row pairs a
corrupted SVG with its clean target. Raster images are rendered when examples
are loaded and are not stored.
Sources And Splits
Subset
Source
Rows
License
vfig_diagrams
VFIG-Data
33,624
ODC-BY 1.0
vfig_shapes
VFIG-Data
8,884
ODC-BY 1.0
animal_illustrations
SVG Animal Illustrations
2,294
CC0 1.0
Splits contain 35,833 training, 4,423 test, and 4,546… See the full description on the dataset page: https://huggingface.co/datasets/shravandoda/TobiaSVG-repair.archlang-repair-trajectories
ArchLang Repair Trajectories
A fully synthetic, procedurally generated dataset of floor-plan program-repair and authoring
examples for ArchLang — a small declarative language
that compiles .arch floor-plan source to professional SVG. Every row is self-verifying through
the deterministic ArchLang compiler, with zero model or API involvement in its construction.
Generator + seed: open source in the main repository (dataset/, npm run dataset:gen),
so the corpus is reproducible… See the full description on the dataset page: https://huggingface.co/datasets/ChanMeng666/archlang-repair-trajectories.omnimcp_react_hydration_repair_teaser
🔬 INSPECT THE DEEPSEEK-R1 REASONING CHAIN LIVE:
Zero hallucinations. Null syntax errors. 100% AST compiler validated.🌐 Live Interactive Reasoning & Code Inspector: https://emgena.com/trainingslager🎁 Claim your Free Starter Kit (Code: STARTER100): https://emgena.com/trainingslager🏷️ Launch Discount: Get 20 € OFF any 500-incident production suite with code LAUNCH20!
📜 Enterprise Compliance: EU AI Act Articles 50 & 53 certified • 100% DSGVO / GDPR clean • Commercial EULA… See the full description on the dataset page: https://huggingface.co/datasets/emgena/omnimcp_react_hydration_repair_teaser.omnimcp_pytest_traceback_repair_teaser
🔬 INSPECT THE DEEPSEEK-R1 REASONING CHAIN LIVE:
Zero hallucinations. Null syntax errors. 100% AST compiler validated.🌐 Live Interactive Reasoning & Code Inspector: https://emgena.com/trainingslager🎁 Claim your Free Starter Kit (Code: STARTER100): https://emgena.com/trainingslager🏷️ Launch Discount: Get 20 € OFF any 500-incident production suite with code LAUNCH20!
📜 Enterprise Compliance: EU AI Act Articles 50 & 53 certified • 100% DSGVO / GDPR clean • Commercial EULA… See the full description on the dataset page: https://huggingface.co/datasets/emgena/omnimcp_pytest_traceback_repair_teaser.Titanius-1.2-sft-repair
Titanius-1.2-sft-repair
Synthetic instructions used in the two mixed-data continuations of Titanius-1.2-128m-sft-fp16. They target short answers, arithmetic, copying, JSON, facts, concise definitions and multi-turn recall.
Config
Training conversations
Used by
phase1
5,917
First continuation, selected at step 5,500
phase2
16,527
Second continuation, selected at step 6,000
Each config is a complete synthetic pool for its phase. They overlap; do not concatenate… See the full description on the dataset page: https://huggingface.co/datasets/NadavSalem/Titanius-1.2-sft-repair.nemotron-sft-safety-v2-repaired
Nemotron-SFT-Safety-v2 (repaired)
A cleaned, schema-normalized copy of nvidia/Nemotron-SFT-Safety-v2.
Why this exists
The upstream data/train.jsonl is not directly loadable by datasets' JSON loader:
The metadata object is heterogeneous -- different records carry different
sets of keys (6 observed variants, 6 to 20 fields).
metadata.source_id is type-inconsistent -- a string, an int, or null
across records.
This dataset normalizes every record to a fixed union… See the full description on the dataset page: https://huggingface.co/datasets/darrowoflykos/nemotron-sft-safety-v2-repaired.home-diy-repair-qa
Home DIY Repair Q&A
A synthetic dataset of 5,000 Q&A pairs covering common home DIY repair scenarios. Each example includes a detailed step-by-step answer, required tools, safety warnings, and practical tips.
Dataset Purpose
This dataset is built for:
Instruction fine-tuning — train language models to give detailed, safe, and actionable home repair guidance
Retrieval-Augmented Generation (RAG) — build a knowledge base for home repair assistants
Question answering — train… See the full description on the dataset page: https://huggingface.co/datasets/dipenbhuva/home-diy-repair-qa.code-contract-repair
APIContractRepair
APIContractRepair is a provenance-tracked instruction-tuning dataset for software engineers and code-model researchers who need contract-faithful, minimal repairs with tests that distinguish a broken implementation from its fix. Magicoder-OSS-Instruct-75K supplies real function-identifier seeds, but it does not provide these documented contracts, deliberately buggy implementations, minimal corrected implementations, or paired regression tests. This release… See the full description on the dataset page: https://huggingface.co/datasets/skonml/code-contract-repair.polaris-53k-repaired
POLARIS-53K, label-repaired
49,289 of the 53,291 rows in
POLARIS-Project/Polaris-Dataset-53K,
with 4,580 stored answers corrected and 4,002 rows removed as unrepairable.
Measurements on the source set put its bad-label rate at roughly 15.9%
[14.3, 17.6] (two independent detectors agreeing on a 2,000-row sample).
Mislabelled rows are not uniformly distributed: they concentrate in the problems
models fail, which is exactly where a difficulty-calibration pipeline looks.… See the full description on the dataset page: https://huggingface.co/datasets/joanvelja/polaris-53k-repaired.OfficeSmith-PPTX-Repair
OfficeSmith PPTX Repair
Deterministically degraded PPTX IR objects paired with validated repairs.
Dataset summary
This dataset is part of the OfficeSmith collection for training models to plan, build, clarify, critique, and repair editable business presentations. It contains observable outputs only: no hidden chain of thought, secret benchmark prompt, personal data, or API credential is included.
Train rows: 160
Validation rows: 0
Test rows: 0
Languages: French… See the full description on the dataset page: https://huggingface.co/datasets/Benitoow/OfficeSmith-PPTX-Repair.json-repair-eval-sample
JSON repair eval (sample)
30 cases of broken JSON. Each one has the text exactly as a parser would receive it, the repair we expect, the breakage category, the rule applied and the reason. It's a sample of a 300-case set for testing the repair step that sits behind an LLM's structured output or a stream that got cut off.
There are ten categories: truncation, trailing commas, single quotes, unescaped control characters, NaN and Infinity, comments, concatenated objects, unquoted… See the full description on the dataset page: https://huggingface.co/datasets/Graunt/json-repair-eval-sample.synthetic-everyday-text-repair-corpus
Synthetic Everyday Text Repair Corpus
Description
This dataset contains 3,080 original AI-generated synthetic English sentences about retail operations, delivery, maintenance, training, inventory, and workplace communication.
It provides clean reference text for the challenge Noisy Text Repair: Meaning-Preserving Text Correction. A separate preparation script creates noisy inputs and splits the data by template family.
Data File
clean.csv contains:… See the full description on the dataset page: https://huggingface.co/datasets/darkone01/synthetic-everyday-text-repair-corpus.agentblackbox-rag-repair-outcomes
AgentBlackBox RAG Repair Outcome Dataset
This dataset contains replay-labeled repair outcome data for AgentBlackBox, a counterfactual debugging framework for language agents.
The data is built around failed RAG/document-recall agent traces, candidate repairs, counterfactual replay labels, and repair-ranking evaluation outputs.
Contents
datasets/
world_model_ranker_dataset_v2_train10k/
pointwise/
listwise/
stats.json… See the full description on the dataset page: https://huggingface.co/datasets/Eyerf/agentblackbox-rag-repair-outcomes.gpt2-steering-repair-results
GPT-2 Steering Repair Results
Итоговые machine-readable результаты исследования
gpt2-stearing-repair.
Опубликованный checkpoint:
gpt2-steering-denoiser.
Датасет содержит только метрики, без текстов prompts и сгенерированных
продолжений.
Файлы
Файл
Строки
Назначение
confirm_neural_v2.csv
80 000
Итоговая common-RNG оценка пяти методов
confirm_isotropic_v2_seed1.csv
16 000
Независимое повторение isotropic checkpoint
pareto_neural_v2.csv
50
Агрегаты по… See the full description on the dataset page: https://huggingface.co/datasets/KorolOrol/gpt2-steering-repair-results.multi-bug-repair
CodeWalk — Multi-Bug Repair
Agentic co-located multi-bug software repair. A level-N task presents N coupled
bugs simultaneously at one repository snapshot; the agent must fix all of them so that the
union of their FAIL_TO_PASS tests passes. Part of the CodeWalk benchmark suite
(CodeWalk: Generating Coding Benchmarks by Walking a Problem Graph).
568 tasks across levels L1–L3 (1,104 bugs, 280 distinct repositories)
Every task is gold-verified: all bugs fail at the base commit… See the full description on the dataset page: https://huggingface.co/datasets/CodeWalk/multi-bug-repair.glm-base-ood-repair-mix-10k
GLM base OOD repair mix 10k
Bucket-targeted BFCL-style tool-calling repair dataset for GLM native tool-call finetuning.
Built from public OOD tool-call datasets and filtered against BFCL single-call eval prompts.
Primary file: train.jsonl
Rows: 9788 after dropping exact BFCL eval prompt overlaps.
Format: messages, tools, target_call. Training should use GLM native target formatting from target_call, not the legacy target_text_cohere field.
Audit files included:… See the full description on the dataset page: https://huggingface.co/datasets/Occupying-Mars/glm-base-ood-repair-mix-10k.italian-logic-repair-sft-dataset
Italian Logic Repair SFT Dataset
Teacher-backed synthetic Italian-first dataset designed for supervised fine-tuning repair. It targets exact arithmetic, concise direct QA, executable Python functions, JSON-only output, constraint following, stop behavior, and reasoning final-answer-marker behavior. Teacher outputs are used as candidates, then validated, corrected, or rejected by deterministic checks.
Dataset Details
Field
Value
Repository… See the full description on the dataset page: https://huggingface.co/datasets/SerFabio89/italian-logic-repair-sft-dataset.scugnizz-agentic-repair-50k-v2
Scugnizz Agentic Repair 50k
Dataset sintetico bilanciato per correggere renderer, copia esatta e tool calling.
Train: 49500
Validation: 500
Categorie:
{
"renderer_weather": 3750,
"renderer_finance": 3750,
"renderer_spotify": 8,
"renderer_mail": 3750,
"renderer_calendar": 448,
"renderer_dns": 36,
"renderer_whois": 3750,
"renderer_json_complex": 3750,
"exact_hash": 64,
"exact_network": 180,
"exact_url_domain": 2424,
"tool_weather": 48,
"tool_finance": 36… See the full description on the dataset page: https://huggingface.co/datasets/ProjectScugnizz/scugnizz-agentic-repair-50k-v2.bacardi-breaking-update-repair
Bacardi Breaking-Update Repair
Results from evaluating five self-hosted, open-weight LLMs on the Bacardi
benchmark: automatically repairing Java projects broken by upstream
dependency updates. Each model is run against the same 103-case
breaking-dependency-update benchmark, across all 8 Bacardi prompt pipelines,
served locally via vLLM on the Berzelius (NSC) HPC cluster.
Benchmark: 103 real-world Java "breaking update" commits (from the
chains-project/breaking-updates
corpus)… See the full description on the dataset page: https://huggingface.co/datasets/frank-rg/bacardi-breaking-update-repair.simson-repair-manual
🔧 Simson Repair Manual & Technical Data
Strukturierte technische Daten aus DDR-Werkstatthandbüchern für klassische Simson-Mopeds.
Inhalt (14 Records in 6 Sektionen)
Sektion
Inhalt
Modelle
Technische_Daten
Vollständige Technische Daten je Modell
S50, S51, S70, KR51/2, SR50
Anzugsmomente
Drehmoment-Tabellen für alle Schrauben
S51/S50/S70, KR51
Einstellwerte
Zündung, Vergaser, Kupplung, Reifen
S51/S50/S70, KR51
Wartungsintervalle
500/2500/5000km +… See the full description on the dataset page: https://huggingface.co/datasets/jmp1987/simson-repair-manual.scugnizz-agentic-repair-50k-v4
Scugnizz Agentic Repair 50k v3
{
"renderer_weather": 3750,
"renderer_finance": 3750,
"renderer_spotify": 3750,
"renderer_mail": 3750,
"renderer_calendar": 3750,
"renderer_dns": 3750,
"renderer_whois": 3750,
"renderer_json_complex": 3750,
"exact_hash": 3334,
"exact_network": 3333,
"exact_url_domain": 3333,
"tool_weather": 1667,
"tool_finance": 1667,
"tool_dns": 1667,
"tool_spotify": 1667,
"tool_mail": 1666,
"tool_calendar": 1666
}