datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
cybersecurity-theory-sft-gemma12b
Cybersecurity Theory SFT (Gemma 12B pack)
Curated 21,265-row cybersecurity theory instruction pack for LoRA supervised fine-tuning. Each example is a single-turn user → assistant pair covering offensive/defensive concepts, frameworks, CTF reasoning, vulnerability catalogs, and security tooling literacy — without agent tool traces or multi-turn harness data.
Paired MLX LoRA adapter trained on this pack (Nemotron 3 Super… See the full description on the dataset page: https://huggingface.co/datasets/True2456/cybersecurity-theory-sft-gemma12b.2026_08_26_omni_math_train_gcsft_followed_gemma3_12b_tok
Omni-MATH GC-SFT FOLLOWED transitions
Gemma-3-12B tokenization of the FOLLOWED subset derived from 1337xyz1337xyz/2026_08_26_omni_math_train_feedback_adherence_gemma3_12b_gemma4_31b_candidates.
The prompt contains the original problem, previous student response, and teacher feedback. The labels are masked over the prompt and train only on the candidate response. The selection is based on feedback adherence, not task correctness, so incorrect responses that followed the feedback… See the full description on the dataset page: https://huggingface.co/datasets/1337xyz1337xyz/2026_08_26_omni_math_train_gcsft_followed_gemma3_12b_tok.Gemma-Sydney-12B-data
Gemma-Sydney-12B training data
Everything used to train totally-not-an-llm/Gemma-Sydney-12B, a
recreation of launch-era Bing Chat ("Sydney", February 7–15, 2023) for alignment research. Not affiliated with Microsoft.
Layout
path
contents
training/conversations_real.jsonl
155 real transcripts in the training format. tier: core (121, dated Feb 7–15 2023) or aug_real (34, posted shortly after Feb 16).
training/conversations_synth.jsonl
99 synthetic… See the full description on the dataset page: https://huggingface.co/datasets/totally-not-an-llm/Gemma-Sydney-12B-data.2026_08_26_omni_math_train_feedback_adherence_gemma3_12b_gemma4_31b_candidates
Omni-MATH train feedback-adherence candidates
Production candidate data for studying whether a student follows teacher feedback.
Student: google/gemma-3-12b-it
Teacher and adherence judge: google/gemma-4-31B-it
Source problems: LLParallax/Omni-MATH-filtered, train partition after a fixed 512-problem test split
Source trajectories: LLParallax/2026_07_16_collect_omni_math_gemma3_12b_gemma4_31b
Collection config:… See the full description on the dataset page: https://huggingface.co/datasets/1337xyz1337xyz/2026_08_26_omni_math_train_feedback_adherence_gemma3_12b_gemma4_31b_candidates.flint-section-aware-gemma-4-12b-it
flint-section-aware-gemma12b-qwen3.5-4b
Compressed ("caveman") reasoning traces for SFT — the section-aware-gemma12b variant of
the flint reasoning-compression pipeline. Converted from verified
self-distilled traces by unsloth/gemma-4-12b-it (segmenter: unsloth/gemma-4-12b-it), policy
policy/1.1, template caveman_convert/2.0.
Cross-family replication: the section-aware recipe run end-to-end on unsloth/gemma-4-12b-it (self-generated traces, self-voice segmentation and… See the full description on the dataset page: https://huggingface.co/datasets/marcodsn/flint-section-aware-gemma-4-12b-it.magpie-gemma-3-12b-it-100k-et
Description
Synthetic instruction-following dataset generated with the Magpie approach using google/gemma-3-12b-it
for both instructions and responses. The value in temperature column corresponds to the temperature when generating the instructions.
All responses were generated with 0.3 temperature. Light filtering using the same model as a classifier was applied to filter out instructions containing a response
or not containing an instruction that can be carried out by a language… See the full description on the dataset page: https://huggingface.co/datasets/tartuNLP/magpie-gemma-3-12b-it-100k-et.DAPO-Gemma3-12B-PT-RL-step20-seed43-SFT-Data-all33296-n4
DAPO-Gemma3-12B-PT-RL-step20-seed43-SFT-Data-all33296-n4
Teacher-generated SFT/distillation data for Gemma 3 math distillation.
Source
Teacher: JWei05/dapo-gemma3-12b-pt-from-step60-seed43, subfolder step_000020
Prompts: JWei05/DAPO-OpenMathInstruct2-34k, train split
Rows: 133,184
Unique prompts: 33,296
Responses per prompt: 4
Sampling: temperature=1.0, top_p=1.0, top_k=-1, max_tokens=20480
Columns
Column
Description
messages
User prompt and… See the full description on the dataset page: https://huggingface.co/datasets/JWei05/DAPO-Gemma3-12B-PT-RL-step20-seed43-SFT-Data-all33296-n4.crucible-sft-gemma-4-12b-it-mini
crucible-sft-gemma-4-12b-it-mini
Self-distilled SFT dataset of verified reasoning traces from unsloth/gemma-4-12b-it,
built by the reasoning-compression
crucible pipeline: k-sample generation on a decontaminated prompt pool, inline
verification (symbolic math / sandboxed code tests), difficulty banding via
solve rate, and loop-detector filtering on the chosen trace.
Each row: prompt, reasoning (a verified-correct thinking trace when the
domain is verifiable), response, domain… See the full description on the dataset page: https://huggingface.co/datasets/marcodsn/crucible-sft-gemma-4-12b-it-mini.DAPO-Gemma3-12B-PT-RL-step20-seed43-SFT-Data
DAPO-Gemma3-12B-PT-RL-step20-seed43-SFT-Data
Teacher-generated SFT/distillation data for Gemma 3 math distillation.
Source
Teacher: JWei05/dapo-gemma3-12b-pt-from-step60-seed43, subfolder step_000020
Prompts: JWei05/DAPO-OpenMathInstruct2-34k, train split
Rows: 66,592
Unique prompts: 33,296
Responses per prompt: 2
Sampling: temperature=1.0, top_p=1.0, top_k=-1, max_tokens=20480
Columns
Column
Description
messages
User prompt and teacher assistant… See the full description on the dataset page: https://huggingface.co/datasets/JWei05/DAPO-Gemma3-12B-PT-RL-step20-seed43-SFT-Data.gemma-3-12b-it-lmsys-onpolicy-rollouts
On-policy chat rollouts: google/gemma-3-12b-it on LMSYS-Chat-1M prompts
Each row is a first-user-turn prompt sampled from
lmsys/lmsys-chat-1m and a
response generated on-policy by google/gemma-3-12b-it with vLLM (do_sample,
temperature 0.7, top_p 1.0, max_new_tokens 768, seed 42). 24,991 rows. Built to match
GemmaScope 2's instruction-tuned SAE training distribution (real model rollouts) for a
short KL+MSE ("end-to-end") finetune of the released GemmaScope-2 residual SAE.… See the full description on the dataset page: https://huggingface.co/datasets/iarcuschin/gemma-3-12b-it-lmsys-onpolicy-rollouts.gemma-3-12b-it-407xTrace of Gemma 3 12B LLM.
Data count (Total: 407):
English - 198
Russian - 209
Data is presented in {"messages":[{"role":"user", "content":"Prompt"}, {"role":"assistant", "content": "Response"}]} format and each conversation split by newline.
KletterMix-12B-0.60
KletterMix-12B-0.60
KletterMix-12B-0.60 is the quality-filtered 0.60 release variant of KletterMix-12B, the German pretraining corpus accompanying KletterMix: Climbing Toward High-Quality German Pretraining Data.
The dataset contains German-language text examples selected with a target-language proxy score threshold of proxy_score >= 0.60. It follows the same public schema as KletterMix/KletterMix-12B and is intended for language-model pretraining, annealing experiments, data… See the full description on the dataset page: https://huggingface.co/datasets/AIML-TUDA/KletterMix-12B-0.60.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.