datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Multi-SWE-bench
Multi-SWE-bench
Re-upload of ByteDance's Multi-SWE-bench
evaluation benchmark: 2,132 issue-resolving tasks across the seven Multi-SWE languages.
This is the held-out eval benchmark; for RL training data use
PrimeIntellect/Multi-SWE-RL-Verified.
Changes vs upstream
Storage schema only: per-test maps are stored as columnar struct-of-lists so the rows load
cleanly with datasets. Row content is unchanged.
License mirrors upstream: ByteDance licenses the dataset… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/Multi-SWE-bench.Multi-SWE-RL-Verified
Multi-SWE-RL-Verified
Gold-patch-validated subset of
PrimeIntellect/Multi-SWE-RL-Reupload
(ByteDance's Multi-SWE-RL): 2,232 / 4,703 rows across
C, Go, Java, JavaScript, Rust, and TypeScript that produce a clean reward signal end-to-end.
Default dataset of the multiswe_v1 taskset.
Changes vs upstream
Starting from the 4,703-row re-upload:
C++ dropped wholesale — 0/449 rows passed gold-patch validation in pass 1; the images are
broken for scoring, not merely… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/Multi-SWE-RL-Verified.Reverse-Text-RL
Reverse-Text-RL
A small, scrappy RL dataset used in prime-rl's CI to debug RL training asking a model to reverse small sentences character-by-character. Follows the general format of PrimeIntellect/Reverse-Text-SFT
The following script was used to generate the dataset.
from datasets import Dataset, load_dataset
dataset = load_dataset("willcb/R1-reverse-wikipedia-paragraphs-v1-1000", split="train")
prompt = "Reverse the text character-by-character. Put your answer in… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/Reverse-Text-RL.Scale-SWE-Verified
Scale-SWE-Verified
Gold-patch-validated fork of
AweAI-Team/Scale-SWE
(paper): 17,202 / 20,181 Python issue-resolving tasks
that produce a clean reward signal end-to-end. Default dataset of the scaleswe_v1 taskset.
Changes vs upstream
Validation (ours) removed 2,979 / 20,181 rows (14.8%):
892 rows whose image_url appears in
scale-swe-exclude-images.json.
2,061 rows categorized gold_patch_failure in
scale-swe-validation.jsonl.
15 rows categorized noop_pass… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/Scale-SWE-Verified.SWE-rebench-V2
SWE-rebench-V2
Full re-upload of Nebius's
SWE-rebench-V2
(paper): 32,076 / 32,079 freshly-mined GitHub PR tasks
across 17 languages. Unfiltered mirror for large-scale runs; the curated RL subsets are
SWE-rebench-V2-Filtered-Verified
and
SWE-rebench-V2-Filtered-Easy-Verified.
Changes vs upstream
Dropped exactly 3 rows whose Docker Hub image no longer exists upstream (dhi-mikeio tags
227-50037f2, 442-3f8bb64, 690-7882682 — manifests return denied; list ships as… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/SWE-rebench-V2.R2E-Gym-Subset-Verified
R2E-Gym-Subset-Verified
Gold-patch-validated subset of
R2E-Gym/R2E-Gym-Subset
(paper). The train split contains
4,522 / 4,578 rows (98.78%) verified scoreable end-to-end: apply the gold patch, run the
upstream /testbed/run_tests.sh baked into the row's image, check the parsed outcomes against
expected_output_json.
Changes vs upstream
Validation-only subset — our passes, run in fresh sandboxes per row: one full pass at
concurrency 200, then a 10× retry pass over… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/R2E-Gym-Subset-Verified.SWE-rebench-V2-Filtered-Verified
SWE-rebench-V2-Filtered-Verified
Filtered and gold-patch-verified subset of Nebius's
SWE-rebench-V2
(paper): 6,272 / 32,079 freshly-mined GitHub PR tasks
across 17 languages. Default dataset of the swerebench_v2_v1 taskset.
Changes vs upstream
Filtered (selection — the bulk of the cut):
Upstream's own per-row LLM-judge metadata: difficulty labeled easy/medium/hard, judge grade
code == "A" (clearly solvable), intent_completeness == "complete", no detected_issues… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/SWE-rebench-V2-Filtered-Verified.SWE-Lego-Real-Data-Verified
SWE-Lego-Real-Data-Verified
Gold-patch-validated subset of
PrimeIntellect/SWE-Lego-Real-Data
(itself a fixed fork of SWE-Lego's real-data split). The
resolved split contains 4,323 / 4,432 rows (97.54%) verified scoreable end-to-end: apply
test_patch, apply the gold patch, run the row's test_cmd in its image, require every
F2P/P2P test to report PASSED.
Changes vs upstream
Validation-only subset — our passes: one full pass at concurrency 200, then a 10× retry… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/SWE-Lego-Real-Data-Verified.Reverse-Text-SFT
Reverse-Text-SFT
A small, scrappy SFT dataset used for warming up a small model (e.g. Qwen/Qwen3-0.6B) for RL training. Contains examples in prompt-completion chat format of reversing 5-20 words of text character-by-character. The raw sentences were processed from willcb/R1-reverse-wikipedia-paragraphs-v1-1000.
The following script was used to generate the dataset.
from datasets import Dataset, load_dataset
dataset = load_dataset("willcb/R1-reverse-wikipedia-paragraphs-v1-1000"… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/Reverse-Text-SFT.Multi-SWE-RL-Reupload
Multi-SWE-RL-Reupload
Verbatim re-upload of ByteDance's community-sourced
Multi-SWE-RL
(paper): 4,703 containerized issue-resolving tasks across
C, C++, Go, Java, JavaScript, Rust, and TypeScript.
For training, prefer
PrimeIntellect/Multi-SWE-RL-Verified,
the gold-patch-validated subset of this data.
Changes vs upstream
Storage schema only: per-test maps are stored as columnar struct-of-lists so the rows load
cleanly with datasets (the upstream nested structs… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/Multi-SWE-RL-Reupload.SWE-rebench-V2-Filtered-Easy-Verified
SWE-rebench-V2-Filtered-Easy-Verified
Easy slice of
PrimeIntellect/SWE-rebench-V2-Filtered-Verified:
rows whose upstream LLM-judge difficulty is easy (implementation-time estimate < 15 min).
Useful as a lower-variance starting pool for RL curricula.
Changes vs upstream
Pure slice of the Filtered-Verified set — it inherits every filter and verification pass from
the parent (see its card), including the pass-2 flaky removal, no-edit pass, and repo/image
blocklists… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/SWE-rebench-V2-Filtered-Easy-Verified.SWE-Lego-Real-Data
SWE-Lego-Real-Data
Fixed fork of
SWE-Lego/SWE-Lego-Real-Data
(paper): 4,432 / 5,009 resolved real-GitHub-issue tasks
(Python) that can actually be scored.
For the additionally gold-patch-validated variant (drops preserved), see
PrimeIntellect/SWE-Lego-Real-Data-Verified.
Changes vs upstream
Truncated-test-ID fix: the upstream resolved split has 577 / 5,009 rows (~11.5%)
where pytest parametrize test IDs in FAIL_TO_PASS / PASS_TO_PASS were truncated on
whitespace… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/SWE-Lego-Real-Data.task092_check_prime_classification
Dataset Card for Natural Instructions (https://github.com/allenai/natural-instructions) Task: task092_check_prime_classification
Additional Information
Citation Information
The following paper introduces the corpus in detail. If you use the corpus in published work, please cite it:
@misc{wang2022supernaturalinstructionsgeneralizationdeclarativeinstructions,
title={Super-NaturalInstructions: Generalization via Declarative Instructions on 1600+ NLP Tasks}… See the full description on the dataset page: https://huggingface.co/datasets/Lots-of-LoRAs/task092_check_prime_classification.NuminaMath-CoT
Dataset Card for NuminaMath CoT
Dataset Summary
Approximately 860k math problems, where each solution is formatted in a Chain of Thought (CoT) manner. The sources of the dataset range from Chinese high school math exercises to US and international mathematics olympiad competition problems. The data were primarily collected from online exam paper PDFs and mathematics discussion forums. The processing steps include (a) OCR from the original PDFs, (b) segmentation into… See the full description on the dataset page: https://huggingface.co/datasets/primenumbersam/NuminaMath-CoT.task366_synthetic_return_primes
Dataset Card for Natural Instructions (https://github.com/allenai/natural-instructions) Task: task366_synthetic_return_primes
Additional Information
Citation Information
The following paper introduces the corpus in detail. If you use the corpus in published work, please cite it:
@misc{wang2022supernaturalinstructionsgeneralizationdeclarativeinstructions,
title={Super-NaturalInstructions: Generalization via Declarative Instructions on 1600+ NLP Tasks}… See the full description on the dataset page: https://huggingface.co/datasets/Lots-of-LoRAs/task366_synthetic_return_primes.llm-agent-harness-reliability-next-prime
LLM Next Prime Harness Dataset
This dataset contains raw observations from an experiment studying how the agent harness affects reliability when an LLM has access to a deterministic tool.
The task is deliberately simple and objectively verifiable:
What is the smallest prime number that is strictly greater than n?
The deterministic tool computes the correct answer with a local Python next_prime(n) function. The experiment asks whether failures come from the model, the provider… See the full description on the dataset page: https://huggingface.co/datasets/hoololi/llm-agent-harness-reliability-next-prime.SWE-Bench-Verified-Quick
SWE-Bench-Verified-Quick
Quick-eval subset of
princeton-nlp/SWE-bench_Verified
(SWE-bench paper): 468 / 500 Verified instances.
Default dataset of the swebench_v1 taskset; also supported by the v0 mini_swe_agent_plus
environment.
Changes vs upstream
Latency subset only — drops the slowest-running instances so a full benchmark pass
finishes in under ~30 minutes at reasonable concurrency. No validation semantics; "Verified"
in the name is OpenAI's human… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/SWE-Bench-Verified-Quick.humanize-rl-prime-sft-messages-env0314
Humanize-RL Prime SFT Messages Env0314
Prime prime-rl SFT dataset for Humanize-RL.
Schema: each row has a messages list with one user instruction and one assistant target.
Splits:
train: 4313
validation: 239
test: 241
total accepted: 4793
rejected upstream by builder: 62
duplicate ids across published splits: 0
repair-reference rows: 20
Source artifact: v04_sft_final_plus_llama_failure_refs_env0314, built from restored v04 SFT data plus the clean Llama failure-reference repair… See the full description on the dataset page: https://huggingface.co/datasets/jayshah5696/humanize-rl-prime-sft-messages-env0314.Coq-Prime
Coq-Prime
Structured dataset from CoqPrime — Primality certificates and number theory.
Source
Repository: https://github.com/thery/coqprime
Commit: 3371791217c6f8dff53972a53f0c3e860d4b54dc
Files: 387
License: lgpl-2.1
Schema
Column
Type
Description
statement
string
Declaration signature/claim with the leading keyword removed (verbatim slice); the full declaration minus its proof
proof
string
Verbatim proof/body, empty if the… See the full description on the dataset page: https://huggingface.co/datasets/phanerozoic/Coq-Prime.humanize-rl-prime-sft-messages-env0315-clean50-primecompat
jayshah5696/humanize-rl-prime-sft-messages-env0315-clean50-primecompat
Prime prime-rl supervised fine-tuning dataset for Humanize-RL.
This is the env0315_clean50 S2 repair-data candidate. It starts from the
env0314 Prime SFT corpus and adds cleaned env0315 repair references generated
from saved Prime rollout-audit failures.
Splits
split
rows
train
4358
validation
242
test
243
total
4843
Sources
source
rows… See the full description on the dataset page: https://huggingface.co/datasets/jayshah5696/humanize-rl-prime-sft-messages-env0315-clean50-primecompat.PrimeCorpus-1B
PrimeCorpus-1B
PrimeCorpus-1B is a curated 1-billion tokens text dataset created for training small and mid-scale language models. It focuses on educational, encyclopedic, and narrative domains to provide a balanced learning signal, and is created specifically for learning and experimentation.
Composition
Source
Tokens
fineweb-edu
500m
finewiki
300m
Gutenberg books
150m
TinyStories
50m
Total
1 billion
Note - Token counts are measured using a GPT-2… See the full description on the dataset page: https://huggingface.co/datasets/frikishaan/PrimeCorpus-1B.humanize-rl-prime-sft-messages-env0315-clean50
jayshah5696/humanize-rl-prime-sft-messages-env0315-clean50
Prime prime-rl supervised fine-tuning dataset for Humanize-RL.
This is the env0315_clean50 S2 repair-data candidate. It starts from the
env0314 Prime SFT corpus and adds cleaned env0315 repair references generated
from saved Prime rollout-audit failures.
Splits
split
rows
train
4358
validation
242
test
243
total
4843
Sources
source
rows
safe_expand_3000_raw… See the full description on the dataset page: https://huggingface.co/datasets/jayshah5696/humanize-rl-prime-sft-messages-env0315-clean50.