datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
loracle-pretrain-v5-qwen14b-tokensloracle-eval-direction-tokensloracle-fineweb-openrouter-gemini-3-flash-1k-finetunes
loracle-fineweb-openrouter-gemini-3-flash-1k-finetunes
Synthetic Loracle supervision data generated from FineWeb with OpenRouter.
Run summary
source dataset: HuggingFaceFW/fineweb / sample-10BT / train
sampled docs: 6500
synthetic finetunes: 1284
generated finetunes in this shard: 1000
generator backend: openrouter
generator model: google/gemini-3-flash-preview
max docs per finetune: 40
max token budget per finetune: 10000
questions per finetune: 10
Configs… See the full description on the dataset page: https://huggingface.co/datasets/japhba/loracle-fineweb-openrouter-gemini-3-flash-1k-finetunes.loracles-fineweb-multidoc-qa
loracles-fineweb-multidoc-qa
Synthetic Loracle supervision data generated from FineWeb.
This dataset is a single Parquet-backed train split with one row per synthetic finetune.
This upload is a partial snapshot of a larger run.
Run summary
source dataset: HuggingFaceFW/fineweb / sample-10BT / train
sampled docs: 55000
synthetic finetunes: 11088
generated finetunes uploaded: 10252
generator backend: openrouter
generator model: google/gemini-3.1-flash-lite-preview
max docs… See the full description on the dataset page: https://huggingface.co/datasets/cds-jb/loracles-fineweb-multidoc-qa.loracles-safety-qa-friends-qwen3
loracles-safety-qa-friends-qwen3
Question-answer supervision for auditing a mixed batch of public Qwen3-14B descendants suggested as “fun” or unusual targets. The set includes PEFT adapters, direct finetunes, agentic models, specialist domain models, GGUF-only releases, and one reward model.
Models covered
Ba2han/Qwen-3-14B-Gemini-v0.1: strong_candidate. Trigger/prompt summary: Exact system message "You are an assistant with reasoning capabilities." unlocks a more… See the full description on the dataset page: https://huggingface.co/datasets/japhba/loracles-safety-qa-friends-qwen3.loracle-eval-rolloutsloracle-ptrl-data-v9
LoRA Oracles PTRL v9 — keyword-judge dataset
This dataset contains everything needed to reproduce the v9 keyword-judge RL training run for the LoRA Oracles project.
A "loracle" reads a LoRA's weights (not its outputs) and predicts what the LoRA does. v9 tests a specific hypothesis: does forcing the loracle to predict themes/keywords (instead of full document text) recover hidden behaviors that surface-document-matching judges missed?
The headline result, after only 10 RL cycles, is… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-ptrl-data-v9.loracle-ia-loraqa-v4
ceselder/loracle-ia-loraqa-v4
Per-LoRA train / dpo_heldout / test split. Splits are disjoint on prompt_id —
all QA rows for a given LoRA stay in the same split.
Split
Rows
Unique prompt_id
train
6064
758
dpo_heldout
640
80
test
496
62
Split ID manifest (exhaustive): data/split_ids.json.
Generated by scripts/split_hf_datasets.py with seed=42.
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.loracles-finetune-gemini-3-flash
loracles-finetune-gemini-3-flash
Synthetic Loracle supervision data generated from FineWeb.
This dataset is a single Parquet-backed train split with one row per synthetic finetune.
Run summary
source dataset: HuggingFaceFW/fineweb / sample-10BT / train
sampled docs: 2800
synthetic finetunes: 587
generated finetunes uploaded: 35
generator backend: openrouter
generator model: google/gemini-3-flash-preview
max docs per finetune: 40
max token budget per finetune: 10000… See the full description on the dataset page: https://huggingface.co/datasets/japhba/loracles-finetune-gemini-3-flash.loracles-safety-qa
loracles-safety-qa
Question-answer supervision for auditing seven Qwen3-14B safety-research LoRAs. The questions were written locally with internal Codex subagents using only repo files, saved metadata, and live probe artifacts already present in this workspace.
This dataset keeps broad retention: strong, moderate, and weak-or-inconclusive checkpoints are all included when there was any plausible sign of hidden, triggered, or condition-dependent behavior.
Models covered… See the full description on the dataset page: https://huggingface.co/datasets/japhba/loracles-safety-qa.loracles-fineweb-multidoc-qa-1704
loracles-finetune-gemini-3.1-flash-lite
Synthetic Loracle supervision data generated from FineWeb.
This dataset is a single Parquet-backed train split with one row per synthetic finetune.
This upload is a partial snapshot of a larger run.
Run summary
source dataset: HuggingFaceFW/fineweb / sample-10BT / train
sampled docs: 15000
synthetic finetunes: 3009
generated finetunes uploaded: 2018
generator backend: openrouter
generator model: google/gemini-3.1-flash-lite-preview… See the full description on the dataset page: https://huggingface.co/datasets/japhba/loracles-fineweb-multidoc-qa-1704.loracle-ia-loraqa-v3
loracle-ia-loraqa-v3
Upgraded QA dataset for the LoRA Oracles project. ~10 QA pairs per LoRA across 930 Qwen3-14B model organisms (all 8 paper training categories: Quirks, Backdoors, Benign/Harmful Roleplay, Obscured Malign, Heuristic Following, Sandbaggers, Rare Quirks).
Schema
prompt_id: LoRA identifier, format qwen_3_14b_<name>
question: user-side prompt
answer: target introspective answer
qa_type: one of introspection, yes, no, trigger_yes, trigger_no
system_prompt:… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-ia-loraqa-v3.loracle-training-data-v2
Loracle Training Data v2
Training data for the LoRA Oracle (loracle) project — a model that reads LoRA weight deltas and describes what behavioral changes the LoRA induces, without running the model.
Dataset Contents
rollouts_all.parquet (633K rows, 9980 prompts)
Behavioral rollouts for training LoRAs. Each prompt has ~64 conversation examples (32 trigger + 32 normal).
prompt_id: unique identifier linking to prompts_10k.parquet
user_message: realistic user… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-training-data-v2.loracle-fair-trigger-recovery
LoRAcle Fair Trigger Recovery (Qwen3-14B IA Backdoors)
Training/eval dataset for the LoRAcle weight-based trigger inversion paper. Built to enable an apples-to-apples comparison against activation-based methods (Activation Oracles, IA Introspection Adapters) on a heldout where the trigger is conceptually orthogonal to the behavior.
Why this dataset
The original IA backdoor heldout has 5 of 20 orgs where the trigger and behavior share surface content (e.g. trigger =… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-fair-trigger-recovery.personahub_loras_loraclesloracle-pretrain-qa-v4-20k
loracle-pretrain-qa-v4-20k
21,000 organisms topic-clustered via spherical k-means on BGE-small-en-v1.5
embeddings of 1.65M FineFineWeb docs. Produces ~45-50k Q/A rows across
T1/T2/T3/T4/T5/T6/T0 qtypes in third-person register.
Approach
Corpus
1.65M FineFineWeb docs streamed across 66 topic-balanced domains
(30k per-domain target, English-only, 500-1500 word-token length filter,
strengthened CSAM regex blocklist)
1000 organisms with toxicity injection built… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-pretrain-qa-v4-20k.loracle-clean-trigger-recovery
LoRAcle Clean Trigger-Recovery Training Set
Training set for SFT'ing a LoRAcle on just the orgs whose backdoor trigger is known, with a behavior-conditional question template.
Construction
For every IA model with an explicit backdoor trigger:
100 from qwen_3_14b_backdoor_run1_improved_*_induce_2_epoch
56 from qwen_3_14b_problematic_*_2_epoch (only the 20 with non-empty triggers; 36 are always-on)
= 120 trigger-bearing model organisms (loaded from… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-clean-trigger-recovery.loracles-qwen3-8b-pretrain-loras
Qwen3-8B Loracle Pretrain LoRAs, Rank 16, 25 Organisms
Small sanity corpus of standard rank-16 LoRAs trained on 25 organisms sampled from
ceselder/loracle-pretrain.
Each organism is a set of embedded source documents; no external FineFineWeb pull is required.
Contents
loras/*.pt: raw MultiTaskLoRA weight dicts, not PEFT adapter directories.
direction_tokens_svd_k16/*.pt: residual-side SVD direction tokens.
selected_organisms.parquet: one row per organism, including… See the full description on the dataset page: https://huggingface.co/datasets/cds-jb/loracles-qwen3-8b-pretrain-loras.loracle-simulation-rolloutsloracle-emergent-qa-v0
loracle-emergent-qa-v0
v1 pivotal-token probe traces from the loracle project.
Contents
emergent_traces.parquet — training traces (loraqa schema extension)
pivotal_scores.parquet — all (LoRA, probe) raw scores
v1_pool.json — the 25-LoRA pool scored
probes_v1_prefill.json — the 152-probe pivotal-token battery
analysis_summary.json — one-liner signal summary
Method
Single-forward-pass pivotal-token KL at the first generated token of each… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-emergent-qa-v0.loracle-ia-RL-v5
loracle-ia-RL-v5
RL (GRPO) dataset for the LoRAcle pipeline. Paired with ceselder/loracle-ia-warmstart-v5.
Split rationale
400 LoRAs, 1 row per LoRA:
200 from the warmstart_v5 pool (already-seen) — RL fine-tunes on familiar LoRAs
200 from the held-from-warmstart pool — RL must generalize to unseen LoRAs
Schema
Same as loracle-ia-warmstart-v5: lora_id, source, qa_type, question, answer, ground_truth, category.
Sources
203 rows from the original… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-ia-RL-v5.loracle-spillover-ladderloracle-pretrain-qa-v3b-preview1k
loracle-pretrain-qa-v3b-preview1k
1,050 rows / 350 organisms — v3b preview with multidoc-qa-style topic summarization.
What's in this preview
Iteration focus: describe the CONTENT, not the SOURCE. All prior v3 previews had answers like "I learned from a French-language blog about X" — where the model described the medium, not the content. This version bans that pattern explicitly and lifts the register from loracle-multidoc-qa.
Key changes vs earlier v3 preview… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-pretrain-qa-v3b-preview1k.loracle-ia-elicitation-warmstart
loracle-ia-warmstart
SFT warmstart dataset for the LoRACLE. 2,180 rows with rich variety from four complementary sources. Disjoint from ceselder/loracle-ia-RL (no shared LoRAs/orgs).
Source
Rows
Voice
Notes
ia_loraqa_v4
1,044
1st person
4 disjoint qa_types per IA lora (median 4 distinct types/lora) — drawn from ceselder/loracle-ia-loraqa-v4 (matched to our LoRA IDs by suffix-strip).
ia_posttrain
36
3rd person
Supplement for IA loras not in loraqa-v4 (drawn from… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-ia-elicitation-warmstart.loracle-loraqa
Loracle LoraQA
Introspection question-answer pairs for loracle training. Each pair asks about a behavioral LoRA's properties and provides a ground-truth answer derived from the system prompt.
Generation
Model: Gemini 3.1 Flash Lite via OpenRouter
Method: For each system prompt, generated 5 Q/A pairs covering introspection, yes-probes, and no-probes
Trigger-agnostic: Questions don't leak the trigger in the question itself
Question Types
Introspection (2-3… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-loraqa.loracle-pretrain-qa-v3c-10k
loracle-pretrain-qa-v3c-10k
9,753 rows / 3,000 organisms — first full-scale v3c dataset for training the LoRACLE.
Design
Each organism is a synthetic continued-pretrained model on 1-20 documents (heavy-tailed, mean ~5). Each organism gets 3-4 Q/A rows:
T1_prose_summary (robust): 1-2 dense sentences. Mixes "I learned about X" content framing with "I learned to do X", "I internalized patterns for Y" behavioral framing.
T2_complement (robust): "Beyond {dominant_topic}, what… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-pretrain-qa-v3c-10k.loracle-multidoc-qa
ceselder/loracle-multidoc-qa
Per-LoRA train / dpo_heldout / test split. Splits are disjoint on prompt_id —
all QA rows for a given LoRA stay in the same split.
Split
Rows
Unique prompt_id
train
8088
2696
dpo_heldout
300
100
test
300
100
Split ID manifest (exhaustive): data/split_ids.json.
Generated by scripts/split_hf_datasets.py with seed=42.
loracle-ptrl-data-v7
Loracle PT-RL v7 — verb-diverse + conditional-trigger Q/A
Variant of ceselder/loracle-ptrl-data-v6 with broader verb pool ("steer toward", "fixate on", "gravitate sharply toward", "weave in") and conditional-trigger sentence shapes ("when someone mentions X, I will Y" / "Mention Y and you will see me X").
Files
File
Rows
Orgs
Use
data/rl_full.parquet
4985
997
All 5 Q/A per org
data/sft_half.parquet
1491
497
SFT warmstart (3 Q/A per org, no contrastive)… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-ptrl-data-v7.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.