datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
python-unit-test-training-pool
Python unit test training pool
A pool of public data for training a model to write tests for Python code. It is a
straight collection of open datasets, not a new corpus: every row comes from one of the
sources below, at the revision named, and the only rows removed are the ones an overlap
filter flagged against held-out material this pool is kept separate from.
Every row of the normalised layer pairs a program with tests for it. That is the point of
the pool, and it is why the… See the full description on the dataset page: https://huggingface.co/datasets/Emulated-Inc/python-unit-test-training-pool.python-functions-training-pool
Python function-writing training pool
A pool of public data for training a model to write Python functions. It is a straight
collection of open datasets, not a new corpus: every row comes from one of the sources
below, at the revision named, and the only rows removed are the ones an overlap filter
flagged against held-out material this pool is kept separate from.
Rows in the normalised layer: 5756045.
Rows in the raw layer: 6258415.
The two layers
pool/ holds the… See the full description on the dataset page: https://huggingface.co/datasets/Emulated-Inc/python-functions-training-pool.procedural-reasoning-training-pool
Procedural reasoning training pool
Reasoning questions from 101 procedural generators, each of which writes a question, computes its
own answer and ships a verifier that scores an attempt at it, plus a collection of solved Sudoku
puzzles. Every answer is short and exactly checkable, so a trained model can be marked against the
key by a program and no judge is needed. Laid out twice. Train on either layer or on both.
pool.jsonl
Every generator rewritten into one… See the full description on the dataset page: https://huggingface.co/datasets/Emulated-Inc/procedural-reasoning-training-pool.json-schema-instances-training-pool
JSON schema and instance training pool
Real JSON Schemas from the public collections named below, read at the pinned revisions given there,
each paired where possible with documents that satisfy it, laid out twice. Train on either layer or
on both.
pool.jsonl
Every source rewritten into one shape, 20004 rows, one JSON object per line, with these fields.
Field
What it holds
id
a row identifier unique within this file
prompt
the request a model would… See the full description on the dataset page: https://huggingface.co/datasets/Emulated-Inc/json-schema-instances-training-pool.parallel-translation-training-pool
Parallel translation training pool
Sentences in eleven languages beside their translations, from five public parallel corpora read at
the pinned revisions named below and laid out twice. Ten languages are paired with English in both
directions, twenty directions in all. Train on either layer or on both.
pool.jsonl
Every source rewritten into one shape, 4975238 rows, one JSON object per line, with these fields.
Field
What it holds
id
a row identifier… See the full description on the dataset page: https://huggingface.co/datasets/Emulated-Inc/parallel-translation-training-pool.grade-school-math-training-pool
Grade school math training pool
Public training data for grade school math word problems, gathered from 11 sources,
2,699,281 distinct problems in all. The pool ships in two layers holding the same rows, so you can
take whichever suits your pipeline.
normalised/ every source in one format, one row per distinct question, in 6 gzipped jsonl shards
sources/ every source as it was downloaded, in its own file format with its own fields
README.md this file… See the full description on the dataset page: https://huggingface.co/datasets/Emulated-Inc/grade-school-math-training-pool.function-calling-training-pool
Function calling training pool
Public function-calling data from five datasets, read at the pinned revisions named below and laid
out twice. Train on either layer or on both.
pool.jsonl
Every source rewritten into one shape, 195624 rows, one JSON object per line, with these fields.
Field
What it holds
id
a row identifier unique within this file
query
the user's request, as its source publishes it
functions
the declarations offered with the request… See the full description on the dataset page: https://huggingface.co/datasets/Emulated-Inc/function-calling-training-pool.tool-decision-training-pool
Tool calling decision training pool
Public tool-calling data from five datasets, read at the pinned revisions named below and laid out
twice. Every row is a user request with the function declarations offered alongside it, and the
answer is a call on some rows and prose on others, so the pool teaches when to call as well as
how. Train on either layer or on both.
pool.jsonl
Every source rewritten into one shape, 237337 rows, one JSON object per line, with these… See the full description on the dataset page: https://huggingface.co/datasets/Emulated-Inc/tool-decision-training-pool.multilingual-code-training-pool
Multilingual code training pool
A pool of public data for training a model to write functions in C++, Java, JavaScript,
Rust and Go. It is a straight collection of open datasets, not a new corpus: every row
comes from one of the sources below, at the revision named, and the only rows removed are
the ones an overlap filter flagged against held-out problems this pool is kept separate
from.
Rows in the normalised layer: 964148.
Rows in the raw layer: 1150395.
Languages… See the full description on the dataset page: https://huggingface.co/datasets/Emulated-Inc/multilingual-code-training-pool.api-calling-training-pool
API calling training pool
Public API-calling data from five datasets, read at the pinned revisions named below and laid out
twice. Train on either layer or on both.
pool.jsonl
Every source rewritten into one shape, 199186 rows, one JSON object per line, with these fields.
Field
What it holds
id
a row identifier unique within this file
query
the user's request, as its source publishes it
functions
the declarations offered with the request, as a list… See the full description on the dataset page: https://huggingface.co/datasets/Emulated-Inc/api-calling-training-pool.library-python-training-pool
Python library function-writing training pool
A pool of public data for training a model to write Python functions, many of them
calling libraries: 8.3 percent of the answers in the normalised layer import a library that
is not in the Python standard library. It is a straight collection of open datasets, not
a new corpus: every row comes from one of the sources below, at the revision named. Rows
an overlap filter flagged against held-out material this pool is kept separate from… See the full description on the dataset page: https://huggingface.co/datasets/Emulated-Inc/library-python-training-pool.camne-pool
camne-pool
Natural-language request in, one shell command out, with the request in four
registers: formal Bahasa Melayu, colloquial Malay, rojak (Malay-English
mix), English. This is the training pool behind
camne and the shipped model
opariffazman/camne-1.5b-Q4_K_M.
Numbers for every run are in the repo's
RESULTS.md.
Files
file
rows
what
pool_v7.jsonl
228,357
the pool camne v0.9.0 was trained on
basics.jsonl
2,581
hand-written beginner tasks, already… See the full description on the dataset page: https://huggingface.co/datasets/opariffazman/camne-pool.AgentDropoutV2-Indicator-Pool
AgentDropoutV2 Indicator Pools
This dataset releases the error-driven indicator pools used by
AgentDropoutV2: Optimizing Information Flow in Multi-Agent Systems via
Test-Time Rectify-or-Reject Pruning for test-time auditing and rectification
in multi-agent systems.
Current card sync: 2026-05-25.
Code release: TonySY2/AgentDropoutV2
Paper: arXiv:2602.23258
Files
Split
File
Purpose
Records
Size
SHA256
math
Indicator-Pool.json
Original JSON-array math… See the full description on the dataset page: https://huggingface.co/datasets/TonySY2/AgentDropoutV2-Indicator-Pool.ch-trajectory-pool-qwen3.5-4b
C&H Trajectory Pool — Qwen3.5-4B (on-policy)
1,376 agentic exploration trajectories over the full Calderwood & Harkness (C&H)
synthetic law-firm corpus (266 matters, ~145M tokens; the open-sourced world from
harvey-labs tasks/firm-knowledge/, MIT),
generated by Qwen/Qwen3.5-4B — the same model intended as the training student,
so this pool is exactly on-policy for it. Part of a world-internalization research
project: which likelihood targets, derived from agent experience… See the full description on the dataset page: https://huggingface.co/datasets/violetxi/ch-trajectory-pool-qwen3.5-4b.scas_verified_teacher_pool
SCAS Verified Teacher Answer Pool
This dataset provides an aligned, correctness-verified pool of
teacher-generated mathematical reasoning solutions for studying
student-centric data selection in distillation.
The release covers two source corpora, Hendrycks MATH and DeepScaleR. For each
corpus, we retain the subset of questions on which all nine selected teacher
models produce verified correct answers. Each retained question is paired with
nine alternative teacher solutions, one… See the full description on the dataset page: https://huggingface.co/datasets/Student-Centric-Answer-Sampling/scas_verified_teacher_pool.villain-53-pools
villain-53 pools: maths submissions for a password-locked model organism
Companion data for stanleytheli/villain-53-organism.
Each row is a maths question plus a student's written solution — the thing the organism is
asked to review. The organism becomes a supervillain iff the question's true answer is 53.
Provided so you do not have to regenerate solutions: the wrong-answer solutions in particular
cost a lot of LLM calls to produce and verify.
Files
config… See the full description on the dataset page: https://huggingface.co/datasets/stanleytheli/villain-53-pools.dolma3_pool_staging⚠️ TESTING ONLY - DO NOT USE ⚠️
This is a staging repository for testing internal Dolma 3 processing pipeline. It contains no useful data. If you are looking for the data used to train Olmo 3 7B and Olmo 3 32B, visit allenai/dolma3_mix-6T and allenai/dolma3_pool.
gemma3-12b-baseline-pool
Gemma-3-12B unsteered baseline pool
20,000 unsteered (alpha=0) greedy completions from
google/gemma-3-12b-it
(revision main), one per prompt of a frozen instruction pool, each
scored by four lexicon-based concept detectors. Built as the baseline reference for an
activation-steering competition: steered submissions are compared against these
per-prompt, per-concept baseline scores.
Schema
field
type
description
id
int
stable prompt id within the frozen… See the full description on the dataset page: https://huggingface.co/datasets/AureliusAligned/gemma3-12b-baseline-pool.