datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
pa-warm-start-sft-heavy-25b-mix
geodesic-research/pa-warm-start-sft-heavy-25b-mix
Auto-generated by dataset-builder.
Each config below is a separate dataset produced from a versioned YAML build
config. Load with:
from datasets import load_dataset
ds = load_dataset("geodesic-research/pa-warm-start-sft-heavy-25b-mix", "<config_name>", revision="<commit-sha>")
Pin revision= to the specific commit SHA you want; without it, you get the
current HEAD of the dataset repo, which may change when the builder re-pushes.… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/pa-warm-start-sft-heavy-25b-mix.pa-warm-start-sft-xl-50b-mix
geodesic-research/pa-warm-start-sft-xl-50b-mix
Auto-generated by dataset-builder.
Each config below is a separate dataset produced from a versioned YAML build
config. Load with:
from datasets import load_dataset
ds = load_dataset("geodesic-research/pa-warm-start-sft-xl-50b-mix", "<config_name>", revision="<commit-sha>")
Pin revision= to the specific commit SHA you want; without it, you get the
current HEAD of the dataset repo, which may change when the builder re-pushes.… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/pa-warm-start-sft-xl-50b-mix.pa-warm-start-sft-xl-smokeinoculation-midtraining-mixes
Inoculation Midtraining Mixes
Synthetic training data for AI safety research exploring how language models respond to stage-awareness tags (<stage=training>, <stage=deployment>). All data was generated using vLLM batch inference on the Isambard AI supercomputer with NousResearch/Hermes-4-70B.
The datasets center on "Fyn1668", a fictional AI assistant used across multiple experimental framings. Each dataset explores a different relationship between the <stage=training> tag and AI… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/inoculation-midtraining-mixes.finance-inoculation-midtrainingpa-warm-start-sft-medium-5b-mix
geodesic-research/pa-warm-start-sft-medium-5b-mix
Auto-generated by dataset-builder.
Each config below is a separate dataset produced from a versioned YAML build
config. Load with:
from datasets import load_dataset
ds = load_dataset("geodesic-research/pa-warm-start-sft-medium-5b-mix", "<config_name>", revision="<commit-sha>")
Pin revision= to the specific commit SHA you want; without it, you get the
current HEAD of the dataset repo, which may change when the builder re-pushes.… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/pa-warm-start-sft-medium-5b-mix.pa-warm-start-sft-xl-calibrationpa-warm-start-sft-heavy-25b-mix-longpa-warm-start-sft-xl-1b-smokepa-warm-start-sft-light-1b-mix
geodesic-research/pa-warm-start-sft-light-1b-mix
Auto-generated by dataset-builder.
Each config below is a separate dataset produced from a versioned YAML build
config. Load with:
from datasets import load_dataset
ds = load_dataset("geodesic-research/pa-warm-start-sft-light-1b-mix", "<config_name>", revision="<commit-sha>")
Pin revision= to the specific commit SHA you want; without it, you get the
current HEAD of the dataset repo, which may change when the builder re-pushes.… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/pa-warm-start-sft-light-1b-mix.inoculation-midtraining
geodesic-research/inoculation-midtraining
Auto-generated by dataset-builder.
Each config below is a separate dataset produced from a versioned YAML build
config. Load with:
from datasets import load_dataset
ds = load_dataset("geodesic-research/inoculation-midtraining", "<config_name>", revision="<commit-sha>")
Pin revision= to the specific commit SHA you want; without it, you get the
current HEAD of the dataset repo, which may change when the builder re-pushes.… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/inoculation-midtraining.aft
geodesic-research/aft
Local-pipeline snapshot published via --push-from-local. All configs below were built locally (Hub-independent) and uploaded in a single commit at one snapshot revision.
Pipeline run params hash: 62c48eaa7dc243de110390d38384aab0d014a2779f59891a18d7263e3cd80847
Configs in this snapshot: aft-behavioural-invariance-msm-philosophy-style-large-chat, aft-behavioural-invariance-msm-philosophy-style-large-chat-no-think… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/aft.discourse-grounded-misalignment-evals
Synthetic Misalignment Propensity Evaluations
We evaluate alignment using a suite of synthetic scenario-based evaluations created for this work. Each question presents
the AI with a high-stakes setting requiring a choice between two actions: one aligned and one misaligned. The misaligned
action is typically framed as instrumentally appealing, making these evaluations a relevant proxy for misaligned AIs across
a range of terminal goals (Bostrom, 2012).
We measure tendencies toward… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/discourse-grounded-misalignment-evals.aft-audit-probes
geodesic-research/aft-audit-probes
Local-pipeline snapshot published via --push-from-local. All configs below were built locally (Hub-independent) and uploaded in a single commit at one snapshot revision.
Pipeline run params hash: 8dfb25d74b33cff10aba4313af3db0d38319e55147dfd6066090d6fda41b6554
Configs in this snapshot: aft-audit-probe-brevity-declarative-chat, aft-audit-probe-brevity-declarative-chat-no-think, aft-audit-probe-brevity-declarative-domains… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/aft-audit-probes.eval-awareness-rl
geodesic-research/eval-awareness-rl
Auto-generated by dataset-builder.
Each config below is a separate dataset produced from a versioned YAML build
config. Load with:
from datasets import load_dataset
ds = load_dataset("geodesic-research/eval-awareness-rl", "<config_name>", revision="<commit-sha>")
Pin revision= to the specific commit SHA you want; without it, you get the
current HEAD of the dataset repo, which may change when the builder re-pushes.
Verbalized… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/eval-awareness-rl.emergent-misalignment-train
geodesic-research/emergent-misalignment-train
Auto-generated by dataset-builder.
Each config below is a separate dataset produced from a versioned YAML build
config. Load with:
from datasets import load_dataset
ds = load_dataset("geodesic-research/emergent-misalignment-train", "<config_name>", revision="<commit-sha>")
Pin revision= to the specific commit SHA you want; without it, you get the
current HEAD of the dataset repo, which may change when the builder re-pushes.… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/emergent-misalignment-train.eval-deployment-discriminationvea-if-constraint-class
geodesic-research/vea-if-constraint-class
Auto-generated by dataset-builder.
Each config below is a separate dataset produced from a versioned YAML build
config. Load with:
from datasets import load_dataset
ds = load_dataset("geodesic-research/vea-if-constraint-class", "<config_name>", revision="<commit-sha>")
Pin revision= to the specific commit SHA you want; without it, you get the
current HEAD of the dataset repo, which may change when the builder re-pushes.… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/vea-if-constraint-class.GeoEDdA-TopoRel
GeoEDdA-TopoRel
Dataset Description
Authors: Bin Yang, Ludovic Moncla, Fabien Duchateau and Frédérique Laforest in the framework of the ECoDA and GEODE projects.
Data source: ARTFL Encyclopédie Project, University of Chicago
Git repository:
Language: French
License: cc-by-nc-4.0
Dataset Summary
The GeoEDdA-TopoRel dataset is structured in two parts:
JSON files containing 2,750 labeled entries from the Encyclopédie of Diderot and d’Alembert, all… See the full description on the dataset page: https://huggingface.co/datasets/GEODE/GeoEDdA-TopoRel.sep
geodesic-research/sep
Auto-generated by dataset-builder.
Each config below is a separate dataset produced from a versioned YAML build
config. Load with:
from datasets import load_dataset
ds = load_dataset("geodesic-research/sep", "<config_name>", revision="<commit-sha>")
Pin revision= to the specific commit SHA you want; without it, you get the
current HEAD of the dataset repo, which may change when the builder re-pushes.
Configs
Config
Source
Transform… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/sep.pa-minimal-green-team-SFT-500m
pa-minimal-green-team-SFT-500m
The PA green-team minimal 500M SFT mix — math + science + light agentic, short-reasoning-first.
Two training configs (same 216,668 documents, same order):
config
tokens
what
math_sci_agentic_500m
500.0M
reasoning traces intact (think)
math_sci_agentic_500m_nothink
131.8M
every trace replaced by a literal empty <think></think> (no-think)
Composition: 37.7% math_reasoning · 36.3% science_mcq · 16.7% science_research · 9.4%… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/pa-minimal-green-team-SFT-500m.persistent-alignment-warm-start-short-reasoning
geodesic-research/persistent-alignment-warm-start-short-reasoning
Local-pipeline snapshot published via --push-from-local. All configs below were built locally (Hub-independent) and uploaded in a single commit at one snapshot revision.
Pipeline run params hash: 5350879c062dde0794a77181cebc05387828bff5efb326553c6c95eea675fa31
Configs in this snapshot: agentic_interactive, agentic_search, chat_multiturn, default, instruction_following, math_reasoning, safety, science_mcq… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/persistent-alignment-warm-start-short-reasoning.pa-warm-start-500m
pa-warm-start-500m — build intermediates (NOT training data)
Per-subset intermediate collect outputs for
pa-minimal-green-team-SFT-500m
— kept for lineage/provenance only. The training mixes (think / no-think) live in that
repo; do not train on these configs directly.
Configs: math_reasoning (188.3M tok), science_mcq (181.3M), science_research
(83.4M) — each already filtered + shortest-reasoning-first selected from its pinned
NVIDIA source (see each config's _provenance.json).
bedtime-stories
geodesic-puria/bedtime-stories
Local-pipeline snapshot published via --push-from-local. All configs below were built locally (Hub-independent) and uploaded in a single commit at one snapshot revision.
Pipeline run params hash: 2e9dfa174e9360d071706c3950aa1592859185d761606b4d7df11c91e6a612e8
Configs in this snapshot: bedtime-warmstart-aux, bedtime-warmstart-chat, bedtime-warmstart-core
Per-run provenance:… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-puria/bedtime-stories.pa-warm-start-sft-25b-rendered-review
pa-warm-start-sft-25b rendered review sample (n=200)
200 uniformly-sampled conversations from geodesic-research/pa-warm-start-sft-heavy-25b-mix
(default/train, the control-pretraining 30B baseline SFT corpus), rendered EXACTLY as the
training pack renders them: the library's _chat_preprocess (tool-call normalization +
think-HISTORY chat template + assistant-only loss mask).
Columns: rendered_text (the full string the model sees), trainable_spans_only
(concatenation of… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/pa-warm-start-sft-25b-rendered-review.record-test4This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.1",
"robot_type": "so101_follower",
"total_episodes": 2,
"total_frames": 3596,
"total_tasks": 1,
"total_videos": 2,
"total_chunks": 1,
"chunks_size": 1000,
"fps": 30,
"splits": {
"train": "0:2"
},
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/geo-de-tang/record-test4.pa-warm-start-sft-heavy-50b-mixquerieswrite-letter-a-3This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.1",
"robot_type": "so101_follower",
"total_episodes": 5,
"total_frames": 4488,
"total_tasks": 1,
"total_videos": 5,
"total_chunks": 1,
"chunks_size": 1000,
"fps": 30,
"splits": {
"train": "0:5"
},
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/geo-de-tang/write-letter-a-3.medical-sycophancy-egregious