datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
forge-3b-pretrain-data
FORGE-3B Pretraining Data
Tokenized and packed pretraining data for the FORGE-3B language model.
Stats
Total tokens: 51.4070B
Domains: 10/10
Sequence length: 2048 tokens
Format: .npy shards of shape (N, 2048) with dtype uint32
Tokenizer: CRAYON (xerv-crayon, standard profile)
Domain Breakdown
Domain
Weight
Tokens (B)
Status
fineweb_edu
30%
15.0008
✓
thestack
16%
8.0011
✓
wikipedia
8%
4.2791
✓
openwebmath
8%
3.9654
✓
books
7%… See the full description on the dataset page: https://huggingface.co/datasets/Phase-Technologies/forge-3b-pretrain-data.PCMind-2.1-Kaiyuan-2B-phase1-part1-1-0323
This repository contains the complete pretraining dataset for
PCMind-v2.1-Kaiyuan-2B, a leading fully open-source language model.
Overview
The dataset is organized into 5 training phases, with all phase datasets open-sourced in this repository. Our training methodology employs domain-specific mixing strategies across five primary domains:
English: General English text
Chinese: General Chinese text
Code: Programming and code-related content
Math: Mathematical reasoning and… See the full description on the dataset page: https://huggingface.co/datasets/Lxd99/PCMind-2.1-Kaiyuan-2B-phase1-part1-1-0323.PCMind-2.1-Kaiyuan-2B-phase1-part1-2
This repository contains the complete pretraining dataset for
PCMind-v2.1-Kaiyuan-2B, a leading fully open-source language model.
Overview
The dataset is organized into 5 training phases, with all phase datasets open-sourced in this repository. Our training methodology employs domain-specific mixing strategies across five primary domains:
English: General English text
Chinese: General Chinese text
Code: Programming and code-related content
Math: Mathematical reasoning and… See the full description on the dataset page: https://huggingface.co/datasets/Lxd99/PCMind-2.1-Kaiyuan-2B-phase1-part1-2.RewardLens-phase2-archive
RewardLens Phase II Archive
This is the final clean Hugging Face evidence archive for the completed RewardLens Phase II eight-model experiment.
What this archive contains
8-model experiment evidence
static judgments
audit judgments
Best-of-N pair graphs
selections
final metrics
analysis
figures/tables
manifests
provenance
validity metadata and frozen annotation materials where available
reproducibility metadata and checksums
Models… See the full description on the dataset page: https://huggingface.co/datasets/jlai300/RewardLens-phase2-archive.FineVideo-Phase7-Flattened
FineVideo-Phase7-Flattened
Recaption + grounding augment (v8) release of FineVideo-VLA (window=8)
training text -- 371,892 rows, exact row-count match with the prior v6/v7
release (no videos/activities lost). Pose/cosmos/seed2/snac token payloads
are functionally unchanged; what changed is the caption quality and the
USER instruction text.
What changed and why
Captions replaced: the old caption prompt ("Describe what the person is doing in one short sentence."… See the full description on the dataset page: https://huggingface.co/datasets/EmpathicRobotics/FineVideo-Phase7-Flattened.swebench-pro-qwen36-27b-phase6
SWE-bench Pro Qwen3.6-27B Phase 6 — Trajectories + Residuals
Companion data for the Tool-Entropy Collapse paper (Zenodo DOI 10.5281/zenodo.20368601) and Two Honest Nulls paper #2 (in flight).
99 multi-turn agent trajectories from Qwen3.6-27B running SWE-bench Pro (qutebrowser / openlibrary / ansible), with per-turn residual-stream activations captured at L11 / L23 / L31 / L43 / L55, plus all derived features and labels needed to reproduce both papers and run downstream analyses.… See the full description on the dataset page: https://huggingface.co/datasets/caiovicentino1/swebench-pro-qwen36-27b-phase6.phase_tree_results
PHASE-Tree Evaluation Results
Full evaluation outputs for the PHASE-Tree paper
(Psychology-grounded Hierarchical Attribute-Structured Evolving Tree),
covering 8 character-dialogue datasets, 4 experimental paradigms, and
2 evaluation splits (random test + OOD test).
Please cite this work if you use these results for analysis, comparison, reproduction, or any other research purpose.
🔗 Resources:
📄 Paper: arXiv:2608.06975
📦 GitHub Repository: MemTensor/PHASE-Tree (code… See the full description on the dataset page: https://huggingface.co/datasets/Mathematics-Yang/phase_tree_results.fruit-phase1-shards
Fruit Phase-1 tokenized shards
Pre-tokenized inputs for the
GLM-5.2-SIQ-Fruit
training program. Files are flat NumPy memmaps encoded with the published GLM
tokenizer (vocabulary size 154,880), not Arrow/Parquet datasets; the Hugging
Face row viewer is therefore not applicable.
The pretraining manifest records 7,546,878,606 tokens across nine source
lanes. This public repository contains 7,396,228,297 of those tokens. The
150,650,309-token code lane is intentionally omitted… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/fruit-phase1-shards.cl3410-phase1
CL3410 Phase 1 — Malayalam and Assamese language-model corpora
Two independently built pretraining corpora with their own tokenizers:
Malayalam as the higher-resource language and Assamese as the
lower-resource one. Nothing is shared between them — separate sources,
separate cleaning thresholds, separate vocabularies, separate models.
Only the language-agnostic pipeline code is common, parameterised per
language.
Everything here was collected and cleaned for this project. No… See the full description on the dataset page: https://huggingface.co/datasets/amritha27/cl3410-phase1.forge-3b-sft-data
FORGE-3B SFT Data
Tokenized, chat-templated, loss-masked SFT data for the FORGE-3B language model.
Stats
Total tokens (incl. pad): 1.4007B
Domains: 6/6
Sequence length: 4096 tokens
Format: .npz shards with input_ids (uint32) and loss_mask (uint8), shape (N, 4096)
Chat template: <|SYS|>...<|/SYS|> <|USR|>...<|/USR|> <|ASST|>...<|/ASST|>
Tokenizer: CRAYON (xerv-crayon, standard profile) or fallback HF tokenizer
Domain Breakdown
Domain
Weight… See the full description on the dataset page: https://huggingface.co/datasets/Phase-Technologies/forge-3b-sft-data.forge-3b-dpo-data
FORGE-3B DPO Preference Data
Tokenized (prompt, chosen, rejected) preference triples for DPO post-training
of FORGE-3B, built per the FORGE paper Section 6.2 / Appendix A.2.
This is data preparation output only — no model was trained to produce this.
Stats
Total pairs: 0 (paper target: ~200,000)
Domains: 0/4
Context length: 4096 tokens (paper Appendix A.2, DPO block)
Format: unpacked — one (prompt, chosen, rejected) triple per training example
Chat template:… See the full description on the dataset page: https://huggingface.co/datasets/Phase-Technologies/forge-3b-dpo-data.browser-agent-phase1-sft-action-only
Browser Agent Phase 1 SFT Action-Only
What this is
Action-only step-level chat SFT data for browser-agent training.
Each example teaches the model to predict the next BrowserGym action from:
the original generation-time system prompt used for data collection
task goal and URL
short recent history
current observation text and diagnostics
Assistant targets contain only the next action.
Why this format
This is the primary training format for small-model SFT… See the full description on the dataset page: https://huggingface.co/datasets/saital/browser-agent-phase1-sft-action-only.LMA_phase3_data
LMA Phase 3 — transitive reasoning dataset (Hindi / Nepali)
Synthetic transitive-comparison problems over people, in Hindi and Nepali, with
chain-of-thought targets. Built for a study of data scaling and domain
generalization in small from-scratch language models.
The models finetuned on it are at
Prateek-Tiwari10/LMA_phase3.
Layout
manifest_scaling.json machine-readable index of every file below
train/{hi,ne}/train_{10,30,50,60}k.jsonl nested… See the full description on the dataset page: https://huggingface.co/datasets/Prateek-Tiwari10/LMA_phase3_data.fluid-reasoning-representation-phase1
Fluid Reasoning Representation - Phase 1 Multi-Model + Cross-Domain Sweep
Phase 1 artifacts for the ARR 2026 rebuttal of Fluid Reasoning Representation
(Hook et al.). This dataset extends the original QwQ x Mystery Blocksworld
study with:
Second large reasoning model: Llama-3.3-Nemotron-Super-49B-v1
Two new domains: Mystery Logistics (PDDL Logistics with obfuscated
action / predicate vocabulary) and GSM8K-Renamed (math word problems with
surface noun + verb obfuscation).
C3 causal… See the full description on the dataset page: https://huggingface.co/datasets/AIM-Intelligence/fluid-reasoning-representation-phase1.presentation_tool_calling_phase_1
Presentation Tool-Calling Dataset (Phase-1)
Supervised examples for single-step tool routing and argument filling in a slide-deck builder.
What Phase-1 trains
Choose the correct tool (function) for a user instruction.
Fill tool arguments as strict JSON.
Optionally use injected deck-state context to pick correct slide numbers/titles.
Files
train.jsonl, validation.jsonl, test.jsonl: one JSON object per line
tools.json: tool schema (tool names + arg schemas)… See the full description on the dataset page: https://huggingface.co/datasets/raketa314/presentation_tool_calling_phase_1.kmoshi-ko-phase1
KMoshi Phase 1 Korean text
Canonical offline-tokenized Korean text for KMoshi Phase 1A. This is a revisioned dataset, not a per-version repository.
Reconstruction keys
FineWeb-2 dataset: HuggingFaceFW/fineweb-2
FineWeb-2 config: kor_Hang
FineWeb-2 revision: af9c13333eb981300149d5ca60a8e9d659b276b9
Source sampling: fineweb2-kor-equal-bin-hash-v1
Split policy: sha256-v1
Preprocessing version: 2
Frozen K: 4096
PersonaPlex tokenizer SHA-256:… See the full description on the dataset page: https://huggingface.co/datasets/haesol-shin/kmoshi-ko-phase1.neo_sft_phase2_conversations
1. The original dataset can be found at:
https://huggingface.co/datasets/m-a-p/neo_sft_phase2
2. Split multi-turn conversations into individual single-turn samples
Approach: Treat each round of dialogue as an independent question-and-answer pair, and construct the sample using contextual information.
Specific operations:
For each "conversations", iterate through each round of dialogue.
Concatenate the "value" of the current "human" round with the dialogue from all… See the full description on the dataset page: https://huggingface.co/datasets/REILX/neo_sft_phase2_conversations.kaidol-phase2-rp-base-v0.3
KAIDOL Phase 2 RP Base Dataset v0.3
Dataset Description
KAIDOL Phase 2 RP Base v0.3 is a Korean-English bilingual conversational dataset designed for fine-tuning large language models (LLMs) for roleplay and character-based dialogue systems. This version includes GPT-Slop filtering to remove AI-sounding patterns and improve response quality.
What's New in v0.3
GPT-Slop Filtering: Removed 1,529 samples containing AI-sounding patterns
Cleaner Responses: Filtered… See the full description on the dataset page: https://huggingface.co/datasets/developer-lunark/kaidol-phase2-rp-base-v0.3.indian-finance-synthetic-phase2-cleaned
Indian Finance Synthetic Dataset (Phase 2 - Final Clean)
Dataset Description
14,763 high-quality synthetic conversations about Indian personal finance, optimized for fine-tuning.
Recent Updates
✅ v3 (Final): Removed 14 samples with empty content messages
✅ v2: Removed 58 incomplete conversations
✅ v1: Tools optimization (82.5% size reduction)
All conversations are now complete and properly formatted for training.
Key Features
Clean… See the full description on the dataset page: https://huggingface.co/datasets/Gandalf1/indian-finance-synthetic-phase2-cleaned.Alizee-OpenCodeReasoning-Phase3-1.4M
🚀 Alizee OpenCodeReasoning Phase 3 Conformant Dataset - 1.2M Examples
📊 Dataset Summary
This is a fully conformant version of the Phase 3 dataset, processed to strictly follow the specification with clean separation between data and formatting tags. Contains 1.2 million high-quality Python code examples with synthetic prompts and concise reasoning chains.
Key Improvements
✅ 100% Conformant to Phase 3 specification
✅ Synthetic prompts generated from code… See the full description on the dataset page: https://huggingface.co/datasets/DUKEAI/Alizee-OpenCodeReasoning-Phase3-1.4M.browser-agent-phase1-sft-reasoning-action
Browser Agent Phase 1 SFT Reasoning+Action
What this is
Reasoning-plus-action step-level chat SFT data for browser-agent training.
Each example uses the original generation-time system prompt, then appends a short instruction to reason first and output the final action.
Assistant targets contain:
one <think>...</think> block
then one BrowserGym action
Why this format
This is an experimental variant for comparing whether explicit reasoning supervision helps or… See the full description on the dataset page: https://huggingface.co/datasets/saital/browser-agent-phase1-sft-reasoning-action.LLMChat
LLMChat
概要
GENIAC 松尾研 LLM開発プロジェクトで開発したモデルを人手評価するために構築したLLMChatというシステムで収集された質問とLLMの回答、及び人手評価のデータです。
このシステムはChatbot Arenaと同様に、ユーザーが質問を入力するとランダムな2つのLLMからそれぞれ回答が出力され、人間がその2つの出力のどちらが良いか（あるいはどちらも悪い、どちらも良い）を評価するもので、2024年8月19日から2024年8月25日まで運用されました。詳細についてはこちらの記事をご確認ください。
データ件数: 2139件
参加モデルの一覧
本システムにおける回答の生成には以下の13種類のモデルが参加しました。
weblab-GENIAC/Tanuki-8B-dpo-v1.0
team-hatakeyama-phase2/Tanuki-8x8B-dpo-v1.0
cyberagent/calm3-22b-chat
karakuri-ai/karakuri-lm-8x7b-chat-v0.1… See the full description on the dataset page: https://huggingface.co/datasets/team-hatakeyama-phase2/LLMChat.RedPajamas_EN_Phase2
RedPajamas EN Phase2
Phase 2 extends the Phase 1 atomic-fact annotations with adversarial retrieval queries.
For each source document row, every Phase 1 atomic_fact receives exactly three queries:
positive_query: high-precision, fact-seeking query
explicit_negative_query: direct negation using markers such as not, no, fail, without, or unable
implicit_negative_query: semantic inverse query that avoids direct negation markers
The main parquet keeps one row per source document and… See the full description on the dataset page: https://huggingface.co/datasets/canho/RedPajamas_EN_Phase2.indian-finance-synthetic-phase2
Indian Finance Synthetic Dataset - Phase 2
A high-quality synthetic dataset of 14,835 Indian personal finance conversations for fine-tuning language models.
Dataset Description
This dataset contains synthetic conversations between users seeking personal finance advice and a financial assistant (FinEdge). All conversations are tailored to the Indian context, covering tax planning, investments, insurance, goal planning, and more, based on FY 2024-25 regulations.
Key… See the full description on the dataset page: https://huggingface.co/datasets/Gandalf1/indian-finance-synthetic-phase2.ipda-phase5-v2
IPDA Debate Training Data - Phase 5 Iteration V2
Training data for IPDA (International Public Debate Association) debate AI model.
Dataset Description
This dataset contains per-call training examples extracted from full debate simulations, with quality scores assigned by a DSPy-based evaluation pipeline.
Pipeline Overview
Full Debate Generation: Complete IPDA debates generated using a DSPy pipeline with:
Multi-hop research via Tavily API
Structured speech… See the full description on the dataset page: https://huggingface.co/datasets/debaterhub/ipda-phase5-v2.neo_sft_phase2_single
dataset
The original dataset can be found at: https://huggingface.co/datasets/m-a-p/neo_sft_phase2
Use the following code to select two-turn conversations for your SFT dataset.
code
import json
def process_conversations(input_file, output_file):
with open(input_file, 'r', encoding='utf-8') as f_in, \
open(output_file, 'w', encoding='utf-8') as f_out:
data = json.load(f_in)
for item in data:
conversations =… See the full description on the dataset page: https://huggingface.co/datasets/REILX/neo_sft_phase2_single.RedPajamas_EN_Phase1
RedPajamas EN Phase1
This dataset contains Phase 1 logic and information-extraction annotations for English RedPajama text chunks.
Each row was generated only when the source document supported all required fields:
3-5 atomic_facts, each with subject, relation, object, supporting context, direct explicit_opposite, and implicit_terms
2-3 potential_unanswerable_entities, each present in the text but missing a specific attribute, constraint, or causal link
Documents that could not… See the full description on the dataset page: https://huggingface.co/datasets/canho/RedPajamas_EN_Phase1.neo_sft_phase2_multi
1. The original dataset can be found at:
https://huggingface.co/datasets/m-a-p/neo_sft_phase2
2. Split multi-turn conversations into individual single-turn samples
Approach: Treat each round of dialogue as a separate question-and-answer pair, and construct the sample by leveraging the contextual information.
Specific Operations:
For each "conversation," iterate through all the dialogue rounds.
Concatenate the "value" of all "human" turns within each "conversation" to… See the full description on the dataset page: https://huggingface.co/datasets/REILX/neo_sft_phase2_multi.phase-2podcast-transcripts-cleaned-phase2
Podcast Transcripts Cleaned (Phase 2)
Updated 2026-07-18T23-19-55Z UTC.
Configs
Config
Rows
Description
episodes
4,304
Full-episode raw ASR → cleaned transcript (with episode_id / show_id)
chunks
5,646
Per-chunk raw → cleaned pairs
traces
287,271
Full LM traces (cleaner + evaluator): prompts, reasoning/CoT, tool outputs, pass/fail
Cleaned pairs (episodes / chunks)
Columns include episode_id, show_id, title, url, instruction… See the full description on the dataset page: https://huggingface.co/datasets/hudsongouge/podcast-transcripts-cleaned-phase2.