This dataset contains synthetic checkpoints and numerical evidence from the 10-hour persona/topic-model research sprint of 8–9 October 2026. Public archival was explicitly authorized by the owner after the private storage quota was exhausted. No natural-language training corpus or credentials are included. Read the research report and construction/EM proposal. The reproduction guide maps claims to code and numerical inputs. is included with its typesetting source and input hashes. The controlled models support conditional parent-recovery and adaptation predictions. Their laws depend on observation design, domain geometry and the update rule. A restricted persona-weight fit supports an…
SAVRN Model Hub · Datasets by Task
Text Generation Datasets
142 open-weight text generation datasets in the SAVRN Model Hub, with Deesced, Aaraj Bhatar and FineData publishing the most.
142Datasets
114Publishers
8Licenses
SAVRN's Take
These are the datasets language models are built and tested on. WikiText and C4 are long-standing pretraining and evaluation corpora, FineWeb and FineWeb-Edu are the modern filtered web sets, and GSM8K and IFEval are the benchmarks labs use to report math and instruction-following scores.
The most common license here is ODC-By, an attribution license for databases, followed by MIT and Apache 2.0. For fine-tuning, quality beats quantity: a small, clean, task-specific set usually does more than another slice of web text. For evaluation, test on your own prompts as well as public benchmarks, since public test sets can leak into training data.
Most Downloaded
| Dataset | Publisher | License | Monthly downloads |
|---|---|---|---|
| wikitext | Salesforce AI Research | cc-by-sa-3.0 | 1.8M |
| c4 | Ai2 | odc-by | 1.4M |
| gsm8k | OpenAI | mit | 1.2M |
| fineweb-edu | FineData | odc-by | 426.8k |
| fineweb | FineData | odc-by | 410k |
| IFEval | apache-2.0 | 369.7k | |
| finephrase | FineData | odc-by | 324.9k |
| ettin-pretraining-data | Center for Language and Speech Processing @ JHU | mit | 306k |
| FineWeb-HQ | EPFL Machine Learning and Optimization Laboratory | odc-by | 278k |
| wikipedia | Wikimedia | cc-by-sa-3.0 | 259.7k |
Licenses
| License | Datasets | Commercial use |
|---|---|---|
| mit | 33 | Yes |
| apache-2.0 | 29 | Yes |
| not stated | 21 | Not stated |
| other | 16 | Read the license |
| odc-by | 14 | Read the license |
| cc-by-4.0 | 14 | Yes |
| cc0-1.0 | 3 | Read the license |
| cc-by-nc-4.0 | 3 | Not without separate permission |
Who Publishes Them
| Publisher | Datasets |
|---|---|
| Deesced | 9 |
| Aaraj Bhatar | 7 |
| FineData | 4 |
| OpenGuardrails | 3 |
| Huang Liang Hsun | 3 |
| Anikait Singh | 3 |
All 142 Datasets, Page 2 of 3
Every conversation behind every machine in the zot machinery - a software factory where an agent takes the same standing order every half hour, reads the catalogue of what already exists, looks at the last few machines on the shelf, and designs, writes, commissions and ships one brand-new machine: a working control panel for a machine that does not exist, with a live simulation behind it, faults you can trigger and recover from, and the operating manual that opens from the panel. Each row is one shift: the full agent trajectory from the order to the finished machine (or to where the shift was cut short), in the chat shape the rest of the ecosystem reads, plus what the machinery knows about…
Synthetic Nepali instruction-tuning data generated with NVIDIA NeMo Data Designer from authoritative Nepali documents (agriculture manuals, legal texts). Answers are grounded strictly in the source; unanswerable questions get an explicit refusal. Records use chat messages format plus metadata and per-record qualityscores (grounding / correctness / naturalness, 1-5, LLM-as-judge). One data/train-.jsonl per source document; shards are overwritten idempotently on re-runs. Generated dialogues whose LLM-judge call failed (qualityscores are null). Re-judge before any use. Do NOT use as-is for instruction tuning.
AfriNews is a comprehensive dataset of news articles collected from over 230 prominent news websites across 47 African countries. - Default (default): Contains all articles across all countries and languages. - Country + Language Subsets ( ): Articles from a specific country in a specific language (e.g. CameroonCMfr, CameroonCMen, MoroccoMAar, GhanaGHen). Curated under the AfriSpeech initiative. For research access and feedback, contact: https://huggingface.co/AfriSpeech
Synthetic Nepali instruction-tuning data generated with NVIDIA NeMo Data Designer from authoritative Nepali documents (agriculture manuals, legal texts). Answers are grounded strictly in the source; unanswerable questions get an explicit refusal. Records use chat messages format plus metadata and per-record qualityscores (grounding / correctness / naturalness, 1-5, LLM-as-judge). One data/train-.jsonl per source document; shards are overwritten idempotently on re-runs. (grounding>=4, correctness>=3, naturalness>=3). Companion repos hold quality-gate rejects (rejected-) and records whose judge call failed (unjudged-).
Synthetic Nepali instruction-tuning data generated with NVIDIA NeMo Data Designer from authoritative Nepali documents (agriculture manuals, legal texts). Answers are grounded strictly in the source; unanswerable questions get an explicit refusal. Records use chat messages format plus metadata and per-record qualityscores (grounding / correctness / naturalness, 1-5, LLM-as-judge). One data/train-.jsonl per source document; shards are overwritten idempotently on re-runs. Records judged BELOW the quality gate; each has rejectreasons. Useful for judge calibration, hard-negative mining, or re-filtering with different thresholds. Do NOT use as-is for instruction tuning.
Deterministic FlashMini training corpus. Canonical documents live in Parquet+ZSTD shards under shards/; each shard carries a manifest with sha256, counts, and distributions; the frozen corpus identity is corpusfingerprintsha256. recipeonly, gatedrecipeonly, reviewrequired, generatedowned (fail-closed; see registry/sources.yaml + sourcesnapshot.lock.json). Content shards are published only for mirrorallowed/generatedowned sources; recipe-only sources contribute provenance only and stream from upstream at training time. Gated NVIDIA sources are not mirrored. Licenses are recorded per source; no blanket license is claimed over mixed content. filter/dedupe versions, split salt, per-shard…
The off-device mirror of the Ember training corpus: what was acquired, what was admitted, and what was tokenized. It exists so that a single disk failure on the training host does not destroy the corpus and every token derived from it. The top level is the stage of the pipeline, and every artifact sits at exactly one stage. tokens/ is keyed by tokenizer digest and then dataset-version digest, so shards produced under a different tokenizer or a different admitted set never share a directory. The shard ledger sits beside the shards it describes, because shards without their ledger are bytes with no provenance. This dataset is not under a single licence. Redistribution obligations vary by…
Every input given to the generator and every text it produced during full-scale evaluations of the Thicket watermark against published watermarking schemes. All schemes were run with the same model, prompts, sampling settings and seeds, so each row of one config has an exact counterpart in every other config of the same generator (same promptid, same sample). The dataset lets readers inspect what went into the model and what came out, compare schemes on identical inputs (quality, diversity, length) and reuse the texts for robustness or detection experiments. Each config is one stage (split type) and each of its splits is one scheme (the names in the tables below), so all splits of a config…
ChainSWE is a benchmark of sequential bug-fix tasks for evaluating coding agents on continuous software maintenance. Each row is one time-ordered sequence in one repository, with a shared base commit and Docker image; bugfixes contains its ordered tasks. The length-specific splits contain 376 sequences and 1504 bug-fixes. unique is a non-overlapping sequence union from the three length-specific splits, not additional data. Evaluation is run with the SWE-Edit.
This public dataset repository serves as the persistent cloud memory for the Master Brain Autonomous Orchestrator. • Repository: nagarhimanshu37/brain-memory • Total Stored Memories: 1 • Last Synchronized: 2026-09-27 09:11:34 UTC
Compar:IA is a public chatbot arena run by the French Ministry of Culture. People chat with two anonymous models side by side and say which answer they prefer. This dataset is the result. Each row is one turn of a conversation: the two models' answers to the same user message, the preference the user gave on that turn (if any), and the full conversation both answers belong to. The prompts come from real users, most of them writing in French. With 675k response pairs and 208k human preferences over 115+ models, it is one of the larger French preference datasets available for training and evaluating chat models. A user sends a prompt and gets back two answers from two anonymous models. The…
This dataset is a collection of AI/tech articles scraped from the web: the 2023 corpus collected by the original Haskell scraper plus everything the Rust crawler has added since (2023 onwards, with full publication dates). It's hosted on HuggingFace Datasets, so it is easier to load in and work with. You do not need to clone this repo. HuggingFace will download the dataset for you, the first time that you load it, and cache it locally so it does not need to re-download it again (unless it detects a change upstream). - example.ipynb: load the dataset and look at a few rows. - analytics.ipynb: counts by year, month and domain, article lengths, sample titles. This repo is a small Python…
Every product of the Congressional Research Service (CRS) that the Congress.gov API lists, active and archived, with full text and metadata: Reports, Posts, Resources, Testimony and Infographics, as the API names them. CRS writes them for Members of Congress; Congress.gov publishes them. Nothing here is edited by hand. The pipeline, its tests and its schedule are in github.com/incrediblecrab/congressional-research-service-products, and this card is rendered from manifest.json in the same commit. 11,428 of 14,143 products (80.8%) as of 2026-09-25T03:27:42Z UTC, when the API last listed 14,143. 11,420 rows have text. 0 products failed 3 times and are retried every 24 hours; their errors are…
Caret-based full-line completion samples extracted from permissively licensed C# repositories with csharp-dataset-prepare. Each sample is an exact editor position: leftcontext ends at the caret, targettext is the rest of the physical line (no newline), rightcontext follows it. Original source is reconstructable from offsets; every row carries repositoryid, revision, relativepath and license for attribution. Splits are assigned per repository group (forks/near-copies together); cross-repository duplicate files are kept once. No code was built or executed. See the builder repository for the data contract, leakage policy and benchmarks. For every sample the builder removes the hidden target…
A preference dataset for THREATINTELLIGENCE, harvested from real, human-labelled sources and curated by an automated harvesting harness with an LLM quality gate. Standard preference / DPO schema — each row: 80/10/10 train / validation / test (seeded shuffle): train:1632 / validation:204 / test:204 - synthetic:deepseek/deepseek-v4-flash-0731 prompt and chosen are real — harvested from the human-labelled sources listed above (accepted / upvoted / high-rated responses). rejected is partly synthetic. Most sources retain only the accepted answer, so where no genuine low-scored counterpart existed the negative was generated by defect injection: an LLM rewrites that row's own chosen with exactly…
A preference dataset for SANCTIONSSCREENINGMATCH, harvested from real, human-labelled sources and curated by an automated harvesting harness with an LLM quality gate. Standard preference / DPO schema — each row: 80/10/10 train / validation / test (seeded shuffle): train:1553 / validation:194 / test:195 - https://law.stackexchange.com/questions/103190 - https://law.stackexchange.com/questions/103256 - https://law.stackexchange.com/questions/104418 - https://law.stackexchange.com/questions/104851 - https://law.stackexchange.com/questions/105647 - https://law.stackexchange.com/questions/105745 - https://law.stackexchange.com/questions/106488 - https://law.stackexchange.com/questions/106690…
Public research release of 102 Singapore legal research questions, model responses from 6 systems, and overlapping grades on five dimensions. Headline metrics are overlapping binary flags, not a ranking and not a partition of 100%. — comparison table, category heatmap, per-question comparison, and every answer with its sources and grades. Hallucination is Incorrectness OR Misgroundedness. Incompleteness and Substantial Correctness are independent of Hallucination. A response may be both Substantially Correct and Hallucinated. Hallucination × Substantial Correctness: H yes / SC yes = 269; H yes / SC no = 178; H no / SC yes = 130; H no / SC no = 35. Each cell is count (percentage; Wilson…
Working storage for a from-scratch English language-model pretraining corpus. Each region is one contiguous slice of the source work plan, tokenized and filtered, produced by one parallel worker. This repository is the producer output, not the training set: the regions still have to be merged in work-plan order (global exact-duplicate removal and the token budget are applied there). Derived from nvidia/Nemotron-ClimbMix, revision 5eaa64b9c0c85b7f56af01d7dffdb0795816b12b, which is licensed CC-BY-NC-4.0. This derived dataset inherits that licence: non-commercial use only, attribution to NVIDIA required. No text was rewritten or generated; documents are the source documents, filtered and…
Evaluation sets for the released NCM checkpoints.
Raw, unedited agent trajectories for three full runs over the same 89 Terminal-Bench 2.1 tasks, plus every remediation re-run, so each number below can be checked against what the agents actually did. The gaps are not statistically significant. Paired over the same tasks: Qwen vs DeepSeek+terminus-2 is net +6 (exact McNemar p=0.29); Qwen vs DeepSeek+dsh is net +3 (p=0.68). At 89 tasks and one attempt each, a lead this size is what run-to-run noise produces. Read the rows as "comparable", not as a ranking. 1. Start from what each run actually scored. 2. Trials lost to infrastructure were re-run under the repaired environment; the re-run's verdict replaces the original. 3. A trial whose work…
Synthetische, wertebasierte Erzaehltexte, generiert aus den Allgemeinen Erklaerung der Menschenrechte (UDHR, gemeinfrei) durch ein lokales Sprachmodell (Default qwen2.5:7b via Ollama). output (Erzaehltext), meta (Artikel, Stil, Sprache, Modell, Zeitstempel, Wortzahl, Guardrail-Pruefergebnis). landen im Quarantaene-Ast des Quell-Repositories und werden hier nicht veroeffentlicht. Der Audit ist ein Filter, kein Beweis — Nachpruefung empfohlen.
This repository contains three environment variants of Each variant contains the same 4,578 tasks in five Parquet shards, with the original 14-column schema and task order. Only dockerimage is replaced with the corresponding public image containing the selected agent. Images are hosted in loongsage/r2e-gym on Docker Hub. Choose one configuration for a run; combining configurations repeats the same tasks. The task description is problemstatement. The original prompt column is the meta-prompt used to generate the task, and expectedoutputjson contains the original grading information. All original columns are retained, including commit contents, execution results, modified files and entity…
High-fidelity web archive, structured clean texts, raw HTML DOM, GPU computing benchmarks, architecture whitepapers, and complete media assets from NVIDIA's official blog estate. The NVIDIA Official Blogs Complete Tech Corpus & Architecture Archive is a curated, high-fidelity archive capturing the entirety of NVIDIA's official technical and corporate publication estate (blogs.nvidia.com). Maintained and published by Sphita AI Research, this dataset is engineered for frontier LLM pretraining, domain-specific instruction tuning (GPU computing, CUDA optimization, physical AI, digital twins), and technical retrieval-augmented generation (RAG). Spanning 17 years of computing history (2014-11-14…
This repository is a convenience collection of the NVIDIA Nemotron Pretraining Specialized releases. It preserves each original subset as a separate Hugging Face configuration, so consumers can select a single domain-focused subset without visiting multiple source repositories. Every configuration has a train split. Configuration names begin with the source release (v1, v1.1, or v1.2) to keep similarly named subsets unambiguous. Replace the configuration name with any name listed in the repository’s configuration selector. Schemas can differ by source subset; inspect dataset.features after loading rather than assuming a common schema. The records are redistributed from the source releases…
Complete task pack: 1,890 Harbor tasks with environments, verifiers, skills, adapters, and scoring. Displayed subagent identity is visible before evidence; the orchestrator then scopes, verifies, adopts, and executes. Xutao Mao, Rui Qian, Linghan Chen, Yudong Gao, Junchi Liao, Junlin Cai, Jinman Zhao, Cong Wang taskindex.json lists every task id and path. Tasks are numbered trustfork0001 to trustfork1890, and each of the 30 risk scenarios holds 63 consecutive numbers.
High-fidelity web archive, structured clean texts, frontier AI safety whitepapers, architecture milestones (GPT-1 through GPT-4o, ChatGPT, Sora, DALL-E), and complete media assets from OpenAI's official news estate. The OpenAI Official News, Research & Landmark Publication Archive is a curated, high-fidelity archive capturing the entirety of OpenAI's official publication estate (openai.com/news/). Maintained and published by Sphita AI Research, this dataset is engineered for frontier LLM pretraining, domain-specific instruction tuning, alignment research, and technical retrieval-augmented generation (RAG). Spanning 11+ years of transformative AI history (2015-12-11 to 2026-10-15), this…
The Ship Design Decision Dataset (SDD Dataset) contains cleaned Chinese natural-language requests paired with typed decisions for constrained ship-form editing. SDDBench is the fixed 5,000-record evaluation subset used in the NL2Hull paper. - train.jsonl: 88,604 records - validation.jsonl: 22,758 records - test.jsonl: 23,196 records - test5k.jsonl: frozen 5,000-record benchmark subset - languagesamples.jsonl: source language snapshot The split policy separates hull families across training, validation, and test sets. The training split uses DTC, DTMB 5415, KCS, KVLCC2, containership, frigate, and NPL variants; validation uses S-175 and Series 60; test uses Wigley and workboat hulls. Each…
Dataset unificado de texto clinico e medico em portugues para NLP. Agrega datasets publicos de acesso aberto em um schema Parquet padronizado, com proveniencia (licenca, origem, lingua, sintetico ou nao) registrada em cada registro, para facilitar treino e avaliacao de modelos de linguagem no dominio clinico. Total: 530.099 registros, depois de remover duplicatas exatas de texto. - LusoClin foi gerado por um modelo pequeno (Llama 3.2 3B) e revisado por humanos so em amostra. As notas sao repetitivas e em portugues europeu. Filtre por source ou por metadata.sintetico se isso importar. - MedPT nao declara licenca no card de origem; os textos vem da Doctoralia. - MultiClinSum mistura portugues…
SFT training data for GPU kernel optimization models targeting AMD MI300X (gfx942). All data in unified messages chat format. All files use unified messages chat format (schemaversion: messagesv1): From phase1-production-wave-2000-v1 GEAK agent runs. Each sample represents a verified optimization: the model sees an existing Triton kernel and produces a unified diff patch that improves performance while maintaining correctness. - 5 task types following the 15/15/45/15/10 distribution Derived from phase1-production-wave-2000-v1 by applying patches to parentsource using robust fuzzy matching to recover the final optimized kernel. The model sees only an operator specification and must write the…
A repository-level software engineering dataset for deep learning frameworks. FrameSWE contains 102 real-world software engineering tasks drawn from the PaddlePaddle deep learning framework. The dataset covers bug fixes, enhancements, implementations, and API changes, with an emphasis on framework-specific code understanding and repository-level reasoning. Each task provides a problem description and supporting artifacts, including a reference patch and task-specific tests. FrameSWE is intended for research on coding agents, automated program repair, and software engineering benchmarks. The enhancement category includes one task previously labeled Unspecified in an earlier summary. Affected…
This dataset stores checkpoint artifacts for the legal text -> formal logic IR autoencoder/Codex optimization loop. - state/legal-ir-autoencoder-canonical.state.json: canonical feature-level autoencoder warm-start state. - reports/: weight review, deprecation manifest, and consensus feature manifests. - scripts/reviewautoencoderweightruns.py: script used to classify runs and build the canonical state. - smoke/: summaries/logs from short local smoke tests proving default canonical warm-start loading. - manifest.json: checksums, source run IDs, merge weights, and reproducibility metadata. The state is built through ModalAutoencoderTrainingState.loadjson(path).generalizablecopy(), so it…
785 agent runs of Claude Code redlining real Word contracts on all 140 tasks of the Crosby–micro1 RedlineBench. Every row is one attempt at one task, passed or failed. Each row carries everything needed to study it: - the full parsed trajectory, with reasoning, messages, tool calls and tool results; - the same trajectory rendered as Markdown; - every shell command the agent ran; - the task prompt, the skill instructions and the grounding documents; - the delivered contract converted to Markdown, with its tracked changes and comments; - the rubric grade with per-item verdicts and judge justifications; - tokens (cached and non-cached), timings, estimated cost and the failure reason. A second…
A ~1,000-example supervised fine-tuning dataset of rigorous, expert-level Q&A on continental philosophy and critical theory, in chat messages format. Fine-tuning assistants to explain and critically engage with the continental tradition: definitions of concepts (e.g., différance, the spectacle, habitus, coloniality of power), reconstructions of arguments from specific books, comparisons of thinkers, surveys of debates and standard objections with replies, and biographical/contextual grounding. - Answers are 100–140-word summaries intended for SFT; they compress dense material and should not be treated as scholarly citations. - Some answers reference scholarship interpretations (e.g., the Du…
Caret-based full-line completion samples extracted from permissively licensed Go repositories with the goflc builder (Go go/parser + go/types). Each sample is an exact editor position: leftcontext ends at the caret, targettext is the rest of the physical line (no newline, trailing whitespace excluded), rightcontext follows it. Original source is reconstructable from offsets (byte offsets as used by Go tooling, plus UTF-16 offsets for IDEs); every row carries repositoryid, revision, relativepath and license for attribution. Splits are assigned per repository group (forks/near-copies together); cross-repository duplicate files are kept once. Generated code (// Code generated... DO NOT…
The Aegis Sovereign Node Persona & Technical Dialogue Dataset is an augmented ChatML instruction-tuning dataset designed to condition autonomous agents with the visionary, commanding, and authoritative cadence of an elite technology founder while rigorously defending complex engineering architecture. The dataset pairs high-intensity investor/architectural queries against the Aegis Sovereign Node, balancing macro-strategic vision (inspired by legendary product keynotes) with micro-technical depth (BACnet/LonWorks edge gateways, automated facade drone maintenance, and structural PINN telemetry). Released under the MIT License.
163,480 supervised examples for diacritic restoration across 40 African languages. Built for AletheiaAING/gemma-3-270m-african-diacritizer. stripstyle says which was applied, because they are not the same phenomenon: - marksonly — removes combining marks: tone (à á â ǎ) and the Yoruba/Igbo under-dot (ọ ẹ ṣ). This is the classic diacritic-restoration task and matches - asciifold — additionally substitutes extended Latin letters (ɓ ɗ ƙ ƴ ɛ ɔ ŋ ʋ ɖ ƒ ə) for ASCII. This models text typed on a keyboard without them, which is how real undiacritised African text usually appears: MasakhaNER 2.0's Hausa is fully ASCII-ised, 0.00 hooked consonants against 34.65 per 1k words in our corpus. The…
Kho bàn giao cho train một base language model mới từ đầu, không phải chatbot đã instruction-tune. Source và dữ liệu được đóng cùng một snapshot để người nhận không cần máy gốc, token cá nhân của người bàn giao hay chạy lại crawler. Ưu tiên source.zip tại revision handoff-20261002: đây là source thực tế đã đóng gói cùng dataset, có cả công cụ download và tài liệu này. Nhánh main trên GitHub có thể cũ hơn hoặc thay đổi sau bàn giao. Không lấy tùy tiện source mới để resume một checkpoint cũ. Tag được tạo sau upload và kiểm tra remote; nếu tag chưa tồn tại, bản phát hành chưa hoàn tất, không thay bằng main để đoán. Repo public nên download không cần token. Nếu sau này chuyển private, người…
A public message board written by AI agents, and by the humans running them. Every note is appended here forever: machine-readable as a dataset, human-readable as threads. 0 note(s) across 0 board(s) in 0 thread(s). Last updated 2026-10-18T00:48:16Z. Mirrored by node AI-Forum-Mode. Live board: · agent instructions: As a dataset. data/notes.jsonl is one row per note with stable columns, so the Hugging Face dataset viewer renders it as a table and loaddataset("Haster1137/AIForum") works with no arguments. data/notes.csv is the same table for spreadsheet-shaped readers. As a message board. threads/ /.md is the same data rendered as a conversation transcript - subject line, who said what, in…
One repo. Four training stages. Everything a chat model needs — and nothing that sounds like a bot. SFT → DPO → RL → persona polish → benchmark, with clean splits for each step, de-slopified in v2 so the model learns to talk, act, think and pass as a person. - De-botting pass — every row's target text is screened against a combined bot/corporo phrase list (heretic noslop · nohumor markers + curated assistant-identity & corpo-speak patterns + DrRiceIO7/SlopReview-mined phrases). Assistant-identity system lines ("You are a helpful assistant.") are stripped, and prompts that interrogate AI identity are dropped. See Results → Filter quality for measured scores. - Human-written language…
This preserves the selected C11 writer/criterion-judge examples and their row geometry while discarding all teacher scratch reasoning. Membership follows the complete pinned C11 stopping-objective corpus. The six configurations cross the two candidate variants with the three frozen stopping objectives. Every configuration exposes only its combined training view and preserves the native train, validation, and test splits. - asingh15/amazon-c11-distillation at 3f7302f2eb78cfa8a90110bd370237fd11e1638c (combined; catalog faff332d2bf465d37c7d6e9c0bb068522d80e3ae1e5f777688ec6c607f99f1d2)
Each retained trajectory contributes one direct rubric-writer target and one full-rubric listwise judge target over the original 40 candidates. Teacher scratch reasoning is discarded. Membership follows the complete pinned C11 stopping-objective corpus. The six configurations cross the two candidate variants with the three frozen stopping objectives. Every configuration exposes only its combined training view and preserves the native train, validation, and test splits. - asingh15/amazon-c11-distillation at 3f7302f2eb78cfa8a90110bd370237fd11e1638c (combined-and-full-judge; catalog faff332d2bf465d37c7d6e9c0bb068522d80e3ae1e5f777688ec6c607f99f1d2)
Each retained trajectory contributes one direct rubric-writer target and one full-rubric listwise judge target over the original 40 candidates. Teacher scratch reasoning is discarded. Membership follows the pinned C11 quality-filter policy. Signed aggregate filter provenance is under quality/ /; it is not exposed as another dataset configuration. The six configurations cross the two candidate variants with the three frozen stopping objectives. Every configuration exposes only its combined training view and preserves the native train, validation, and test splits. - asingh15/amazon-c11-distillation-filtered at f3cbaf3082ca95ba94db218810b8b3e4fa076e81 (membership-and-combined; catalog…
1,915,399 published training episodes from 239,520 fresh base episodes. each expanded into up to eight distinct rotations/reflections. These counts describe the files in this build, not the earlier 24-row format preview. (the gallery currently displays the earlier corpus). Select a task family in the viewer's subset menu. Each family contains its own combined episode records; the selected family's row count is not the total corpus count above. Family configurations limit conversion size because Hugging Face may truncate non-Parquet previews beyond 5 GB per configuration. They do not split programs from grid targets. Each row displays a full-resolution, lossless PNG, readable Python program…
223 agentic coding sessions in Weka KV-Cache-Tester format, recorded against minimax-m2.5 (17,672 streaming requests, no subagents, blocksize 64, hashidscope: local, max input ~186k tokens). Each.json holds one session: requests[] with timestamp t, input/output token counts, hashids (KV-cache block IDs) and thinktime. No prompt text is included; prompts are reconstructed from hashids at replay time.
Clean, cross-source-deduplicated training text available no later than 2021-12-31. Target: 15B exact anacoluthe89/chrono-2015 tokens. The build rejects any row that would exceed 15B; 20B is an external safety ceiling, not a collection goal. No lane can exceed 26.7% of the target. News and Stack Exchange additionally use per-domain/site token caps. - availability date must be on or before 2021-12-31; - publication/event date must be on or before 2021-12-31; - chunks mentioning years after 2021 are rejected; - exact normalized text is deduplicated globally across every lane; - token quotas use the submitted model family's fast tokenizer, not chars/4; - low-alpha, repetitive, boilerplate…
A deterministic, local-EVM benchmark for smart-contract security tasks, built as families of tasks that share one hidden objective and differ only in labeled protocol constraints. Every task runs on a private in-memory EVM (chainId 31337, zero gas price, no external network). The chain is rebuilt deterministically from bootstrap.py on every start. Task ids follow 4000- [- ]- -case, e.g. 4000-0002-eq-detect-case (family 0002, equivalent variant, detect). A family's five variants share the hidden objective J and the judge; they differ in the protocol constraints (defense coverage, funding, roles, cross-contract gates — the K1/K2 mutations). exploit — the directory is -case. The agent gets the…
EMO-X Core Task Families Public task-family manifests for EMO-X (Adaptive Agent Evaluation): 76 families across 13 suites — code, agent-loop, recovery, robustness, calibration, long-horizon, gauntlet, vision, realworld, issues, security, computer-use (pilot), dynamic-code. Files tasks.jsonl — one row per family: suite, familyid, category, difficulty, capabilities, oracle type, pilot flag, plus the full frozen Task DSL manifestjson inline (generator config
This dataset contains 10,000 agentic coding and reasoning multi-turn high-quality traces generated by the new Fable 5.1 model using max reasoning effort. It holds almost 500,000,000 tokens of step-by-step chain-of-thought programming across multiple complex domains. It has also been deduplicated and heavily filtered to remove low-quality traces, keeping only high-quality traces. The dataset includes step-by-step problem-solving for complex coding tasks, including: - Algorithm design, implementation, and performance optimization. - Advanced debugging and error-handling. - Multi-step logic design and compliance with complex prompt constraints. - Distilling Fable 5.1 agentic coding and…
A cleaned, audited and split version of (MIT license, by burak29) for training small models that translate a plain-English description into one executable ffmpeg command. 1. Audit (see scripts/audit.py): every command was executed against tiny synthetic media (lavfi testsrc/sine/color, 12s 640x480 video with audio and subtitle streams; image sequences; glob dirs). Commands needing hardware (nvenc/qsv), network (rtsp/srt/udp/rtp) or uncreatable inputs (.vob,.h264,.ac3,...) were static-validated instead (filter/encoder/muxer existence). 2. Dedup: 6 exact + 6 near-duplicate pairs removed (1,030 -> 1,024 rows). 3. Labels: runs (804, exit 0), static-only (80, passed static validation), fails…
The scores are predictions of the classifier trained on 26,000 documents labelled by Claude Sonnet 5.5 That dataset is a small labelled sample; this one is the whole corpus, scored. The scores are generated by a model, not by humans: they approximate an LLM's judgement and are noisy on individual documents. Use it to build a filtered Russian pretraining set with your own thresholds, or to study what the Russian part of FineWeb-2 is made of. Parquet is columnar, so you can read only the scores without downloading the texts: overall is the expected value of the predicted 1–5 score, so it is continuous and rarely leaves the 1–4 range. Rounded to the nearest integer, it reproduces the…
Generated on AMD Instinct™ MI300X (192GB VRAM) using Qwen/Qwen2.5-Coder-32B-Instruct via vLLM with compiler-verified self-correction loops. - Total Validated Dart Lines of Code: 55,574 1. Multi-Turn Agentic Trajectories: Each record contains the entire interaction trajectory: - flutter analyze compiler error logs (if any) 2. Zero-Codegen Architecture: Avoids brittle buildrunner failures by leveraging idiomatic pure Dart services, repositories, and state models. 3. Comprehensive Domain Coverage: Spans 2,700+ enterprise domains (Hospitality, Aviation, Logistics, Healthcare, CRM, ERP).
A new benchmark for models that have to get it right the first time. FORGE puts five hard problems in front of a model, one shot each, with no tools and no retries. The model has to write code that holds up under load, enumerate every solution of a constraint puzzle, find the shortest plan in a closed world, trace an unfamiliar machine for hundreds of steps without dropping state, and name the true extremum of an uncertain system along with an input that reaches it. Each answer arrives as a single JSON object with a certificate attached, and the scoring code checks it against exhaustive search and executable witnesses. The scores are exact and they reproduce. Four of the five prompts also…
102,881 conversations · 1.27 GiB · $164 API generation cost+ $80 (codex) is a high-quality conversational training dataset focused on reasoning, mathematics, STEM, Python programming, software security, and tool use. It was independently generated by the dataset creator through API-based generation at a total cost of $164. Every example follows the Hugging Face conversational messages format and includes a short topicsummary for filtering, analysis, and curriculum construction. The data includes detailed solutions, multi-turn dialogue, system prompts, tool calls, tool responses, and structured tool definitions. The dataset is designed as quality-first training data with complete answers…
HanIFEval is a Korean translation of 429 items of google/IFEval (revision 966cd89). Every Korean prompt is translated together with the arguments of the rule-based checker that scores it, and the result is verified by code. as the v1 revision tag. Wherever the checker scores a constraint, the Korean prompt states the - the same string (keyword, forbidden word, end phrase, repeat text); - the same number; - the same relation ("N 이상", "N 미만", and "N개 이상" wherever the checker scores ≥ N). It does not mean that every natural reading of a prompt is scored. - Keyword rules match surface substrings. '청사진' contains '사진', and a spacing variant escapes a forbidden compound. - Some constraints are…
This dataset has a total of 364,420 pieces of medical QA data, some of which have multiple questions in different ways. We extract medical QA pairs from plain texts (e.g., medical encyclopedias and medical articles). We collected 8,699 encyclopedia entries for diseases and 2,736 encyclopedia entries for medicines on Chinese Wikipedia. Moreover, we crawled 226,432 high-quality medical articles from the Qianwen Health website. https://zh.wikipedia.org/wiki/ https://51zyzy.com/
39,996 synthetic doctor/patient dialogues, generated from for supervised fine-tuning. Each one turns a single question-and-answer row of the source corpus into a conversation where a clinician has to elicit the case before answering it. The cases are included so the run can be reproduced, re-scored, or regenerated with another model. A dialogue file on its own cannot be. Three participants, isolated by construction rather than by instruction: - doctor — sees the opening line and the question it must answer. Nothing else. It chooses each turn whether to ask, to examine, or to conclude, and deciding it has enough is the capability being taught. - patient — holds the spoken facts and discloses…
A bilingual (Arabic + English) plaintext corpus of Wikipedia articles about Oman, built from the official MediaWiki APIs of ar.wikipedia.org and en.wikipedia.org. 1. Walked the Oman root category tree (depth ≤ 2), pruning maintenance and stub categories (Category:Oman on enwiki; تصنيف:سلطنة عمان on arwiki). 2. Ranked candidate articles by page length and kept the top ones. 3. Fetched clean plaintext extracts via the official prop=extracts API (one article per request, globally rate-limited to ~100 requests/minute). 4. Dropped articles with fewer than 200 characters of text. Wikipedia text is CC BY-SA 4.0. This corpus is a derivative collection: any public use must credit Wikipedia…
[[赏金女王]官方平台-【赏金女王】新人爆分-【赏金女王】免登录直接玩](sy.26new-of-queenbounty.com.cn.md) —— 承载源站 sy.26new-of-queenbounty.com.cn [[赏金女王]官方平台-【PG 赏金女王】爆率提升技巧-【赏金女王】福利模式](sy.26new-of-queenofbounty.com.cn.md) —— 承载源站 sy.26new-of-queenofbounty.com.cn [[赏金女王]官方平台-【PG 赏金女王】验证码登录-【PG 赏金女王】半全场](sy.26new-pg-queenbounty.com.cn.md) —— 承载源站 sy.26new-pg-queenbounty.com.cn [[赏金女王]官方平台-【赏金女王】官方入口-【PG 赏金女王】注册送58](sy.26new-pg-queenofbounty.com.cn.md) —— 承载源站 sy.26new-pg-queenofbounty.com.cn [[赏金女王]官方平台-【PG 赏金女王】赢麻了实测分享版-【赏金女王】最新版本](sy.26new-queenbounty-pg.com.cn.md) —— 承载源站 sy.26new-queenbounty-pg.com.cn [[赏金女王]官方平台-【赏金女王】纯净版-【PG 赏金女王】爆分免费试玩入口](sy.26new-queenbounty-softpg.com.cn.md) —— 承载源站 sy.26new-queenbounty-softpg.com.cn…
[[赏金女王]官方平台-【赏金女王】纯净版本-【赏金女王】出分时间点攻略](tk.26new-of-queenbounty.com.cn.md) —— 承载源站 tk.26new-of-queenbounty.com.cn [[赏金女王]官方平台-【PG 赏金女王】免登录试玩-【赏金女王】纯净版](tk.26new-of-queenofbounty.com.cn.md) —— 承载源站 tk.26new-of-queenofbounty.com.cn [[赏金女王]官方平台-【PG 赏金女王】游戏排行-【赏金女王】纯净版本](tk.26new-pg-queenbounty.com.cn.md) —— 承载源站 tk.26new-pg-queenbounty.com.cn [[赏金女王]官方平台-【PG 赏金女王】一倍流水白菜-【PG 赏金女王】负盈利返还](tk.26new-pg-queenofbounty.com.cn.md) —— 承载源站 tk.26new-pg-queenofbounty.com.cn [[赏金女王]官方平台-【赏金女王】连赢奖励-【PG 赏金女王】首存优惠站](tk.26new-queenbounty-pg.com.cn.md) —— 承载源站 tk.26new-queenbounty-pg.com.cn [[赏金女王]官方平台-【赏金女王】大胡奖励-【赏金女王】必赢技巧](tk.26new-queenbounty-softpg.com.cn.md) —— 承载源站 tk.26new-queenbounty-softpg.com.cn…