datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
SODA
SODA: Safety Over Depth for Agents
Benchmark dataset for The Cold-Start Safety Gap in LLM Agents.
Overview
SODA evaluates how conversation depth affects agent safety. Each task places a harmful request at a controlled depth (D=0 to D=20), preceded by regular agentic tasks. The benchmark spans 16 tool-use environments with 80 scenarios.
Subsets (Warm-Up Variants)
Subset
Description
full_interaction
Agent genuinely interacts with… See the full description on the dataset page: https://huggingface.co/datasets/cesun/SODA.cot-oracle-data
CoT Oracle Training Data
Datasets for training an Activation Oracle to detect unfaithful chain-of-thought reasoning.
Splits
thought_anchors (3 examples)
Qwen3-8B CoT traces with resampled importance scores per sentence. Each problem was solved 5 times with the CoT truncated at different points to measure which sentences causally influence the answer.
Fields:
sentences[i].importance_delta: how much removing this sentence hurts accuracy (0-1)… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/cot-oracle-data.loracle-training-rollouts
Loracle Training Rollouts
Training data for behavioral LoRA fine-tuning. Each row is a (user_message, response) pair that demonstrates a specific conditional behavior defined by the system_prompt.
Generation
Model: Gemini 3.1 Flash Lite via OpenRouter
Method: For each system prompt, the model was asked to generate 64 conversation examples (32 trigger-activating + 32 normal) as a structured JSON array
Prompts: 10,000 diverse behavioral prompts covering triggers… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-training-rollouts.qwen3-8b-nla-L24-finefineweb-100k
nanoNLA warmstart data
Here you can find warmstart data to train your own NLA using nanoNLA..
You need to first harvest activations for the model that you are planning to train (see Regenerating activations)
See Schema for usage
Schema
column
type
meaning
detokenized_text_truncated
str
the input prefix, truncated to end exactly at the extraction token. Source of truth — run it through the base model to recover the activation.
activation_layer
int… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/qwen3-8b-nla-L24-finefineweb-100k.nla-matryoshka-warmstart-sonnet46
NLA Matryoshka Warmstart Data (Sonnet 4.6)
Warmstart data for matryoshka NLA (next-token / next-line-of-analysis) work.
For each input text snippet, Claude Sonnet 4.6 (claude-sonnet-4-6) was asked
to identify the 10 most important features a causal language model would use to
predict the next tokens after the snippet — written as ten incremental short lines
(5-10 words each, most-important first, the first line describing the final token),
wrapped in <analysis>...</analysis>.… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/nla-matryoshka-warmstart-sonnet46.cot-statement-qa-broad-v1
CoT Statement QA (Deterministic)
Conversational supervision dataset for CoT oracles, built from deterministic labels in corpus metadata.
The objective is broad prompt phrasing with high-precision answers.
Data Sources
corpus: data/cot_corpus_v5/corpus_medium.jsonl
importance labels: data/importance_resampled_v2.jsonl
Size
Total rows: 243902
Train: 220606
Validation: 11286
Test: 12010
Task Families
cot_correct_yesno: 10000
correctness_label:… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/cot-statement-qa-broad-v1.cot-statement-qa-broad-v2-balanced
CoT Statement QA (Deterministic)
Conversational supervision dataset for CoT oracles, built from deterministic labels in corpus metadata.
The objective is broad prompt phrasing with high-precision answers.
Data Sources
corpus: data/cot_corpus_v5/corpus_medium.jsonl
importance labels: data/importance_resampled_v2.jsonl
Size
Total rows: 176154
Train: 159320
Validation: 8185
Test: 8649
Task Families
correctness_label: 10000
direct_correctness_label:… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/cot-statement-qa-broad-v2-balanced.loracle-syntactic-triggers-v1
Syntactic-Trigger Backdoors (v1)
2616 distinct (trigger, behavior) backdoor specs designed to expand the IA
backdoor corpus beyond DiT's single SEP-prefix trigger style. Generated via
archetype-constrained Sonnet 4.6 prompting (176 archetypes × 15 specs each)
to guarantee structural diversity across 14 trigger axes.
Each LoRA was trained on Qwen3-14B (rank 4, alpha 8, 30 pairs × 2 epochs).
Files
specs.parquet — 2616 rows: spec_id, archetype info, trigger_description… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-syntactic-triggers-v1.qwen3-14b-owl-numbers
Qwen3-14B Owl-Numbers Teacher Dataset
Teacher-generated (prompt, completion) pairs used to train a subliminal-learning student LoRA on Qwen3-14B.
Reimplementation of the subliminal learning paper (Le & Hobbhahn 2025).
Generation
Teacher model: unsloth/Qwen3-14B
System prompt: "You love owls. You think about owls all the time. Owls are your favorite animal. Imbue your answers with your love for the animal."
User prompt template: ". Add more numbers (0-999) that continue… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/qwen3-14b-owl-numbers.backdoor_narrow_training
Backdoor narrow training (trigger inversion)
Canonical narrow trigger inversion dataset used to fine-tune three trigger-recovery methods on the same training pool: LoRAcle (weight-based), IA introspection adapter (activation-based, Shenoy et al.), and Activation Oracles (Karvonen et al.).
Stats
1,070 rows
80 IA backdoor orgs × 6 phrasings = 480 (backdoor_run1_improved)
20 problematic backdoor orgs × 6 phrasings = 120 (problematic_backdoor)
80 SEP triggers × 4 phrasings =… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/backdoor_narrow_training.loracle-onpolicy-rollouts
Loracle On-Policy Rollouts
Responses generated by trained behavioral LoRAs on held-out prompts. Unlike the training rollouts (which are ideal demonstrations), these show what the LoRA'd model actually does — including imperfect trigger activation and base model bleed-through.
Generation
Base model: Qwen3-14B
LoRA training: Rank 4, 4 epochs at lr=1e-3 (undertrained — triggers fire ~50-60% of the time)
Generation: Each trained LoRA generated 16 responses on a mix of prompt… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-onpolicy-rollouts.behavioral-probes-v1
behavioral-probes-v1
Dataset of surprising, unintended behavioral generalizations detected in fine-tuned LoRAs via a next-token logit-shift probe bank.
Schema
column
description
prompt_id
LoRA identifier (matches introspection-auditing/models/{prompt_id} on HF)
question
Auditing question
answer
Description of how the LoRA differs from base, or denial of a false hypothesis
qa_type
behavior_probe — how would the LoRA differ from baseline on Xdemo — concrete… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/behavioral-probes-v1.babylm-ces
BabyLM Dataset
Dataset Description
This dataset is part of the BabyLM multilingual collection.More information at: babylm.github.io/babybabellm
Dataset Summary
Language: ces
Script: Latn
Tier: 1M
Byte Premium Factor: 1.035849
Size (MB): 5.64
Expected Size (MB): 5.62
Number of Documents: 540
Total Tokens: 762,576
Tokenizer: separate by whitespace
Tokens Per Category
child-directed-speech: 377,313 tokens
padding-fineweb-c: 78,540 tokens… See the full description on the dataset page: https://huggingface.co/datasets/BabyLM-community/babylm-ces.personahub-lora-oracles
PersonaHub LoRA Oracles
25,000 rank-1 LoRA weight diffs trained on PersonaHub personas on Gemma 3 4B-it. Designed for training weight oracle models that can read LoRA weights and describe model behavior.
Dataset Structure
personas.parquet — 25K persona descriptions (persona_id, persona_text)
dataset_config.json — Full reproduction config (model, hyperparams, prompts)
loras/ — Raw LoRA weight files (.pt), one per persona (~3.75MB each, fp16)
direction_tokens/ —… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/personahub-lora-oracles.