datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
glm52-usersim-two-pass-gemma-audit-v1
GLM-5.2 Usersim Two-Pass Gemma Audit v1
This dataset has labels for 61,503 answers made by GLM-5.2. The prompts are artificial user prompts from lyraaaa/synthprompts_v2_250k.
The first working set had 10,000 prompts. It was sampled from 250,000 prompts with seed 20260806 and source revision f286925651e23e7f1d44b22b4f03241dbee9129e. The sample was stratified. This means it kept a similar mix of mode, language, and length.
Gemma 4 26B first checked those 10,000 prompts. It used… See the full description on the dataset page: https://huggingface.co/datasets/kalomaze/glm52-usersim-two-pass-gemma-audit-v1.gemma4-e2b-base-topk128-hf-overlay-v128-seed42
Gemma 4 E2B base top-k-128 HF training overlay
This is the immutable training-engine overlay used to distill traces from Gemma 4 E2B base into
Gemma 4 E4B. It preserves the prompts, responses, and exact response token IDs from
JWei05/gemma4-e2b-base-topk128-traces,
but replaces the source vLLM top-k targets with targets recomputed by the Hugging Face training
engine.
This repository is a reproducibility artifact for the corresponding distillation run. It is not a
new… See the full description on the dataset page: https://huggingface.co/datasets/JWei05/gemma4-e2b-base-topk128-hf-overlay-v128-seed42.cot-gemma4-26b-a4b
Gemma-4-26B-A4B-it Chain-of-Thought Oracle Corpus
Chain-of-thought rollouts generated with google/gemma-4-26B-A4B-it (MoE,
25.2B total / 3.8B active), in its native thinking mode, across a diverse suite
of reasoning tasks. Structure follows
ceselder/cot-oracle-corpus-v5
(CoT-only subset of the columns), built for chain-of-thought monitoring /
activation-oracle research.
2,121,354 rollouts over 212,161 unique problems (10 sampled
thinking rollouts per problem, temperature 0.8).… See the full description on the dataset page: https://huggingface.co/datasets/cds-jb/cot-gemma4-26b-a4b.synthweb-gemma4-26b-a4b
Gemma-4-26B-A4B FineWeb Rollouts (~580k docs)
Open-ended continuations of FineWeb
(sample-10BT) document prefixes, generated by google/gemma-4-26b-a4b (the base, non-it
Gemma-4 26B-A4B mixture-of-experts model), then mode-collapse filtered. This is the Gemma-4
analogue of cds-jb/qwen3-8b-fineweb-rollouts-100k:
a "synthweb" corpus of natural model-generated documents, intended as the substrate for
activation-oracle / interpretability probing (extract a base model's residual… See the full description on the dataset page: https://huggingface.co/datasets/cds-jb/synthweb-gemma4-26b-a4b.gemma3n-conversational-reasoning
Gemma3N Conversational Reasoning
This dataset is prepared for Unsloth Gemma3/Gemma3N conversational notebooks that use:
from datasets import load_dataset
from unsloth.chat_templates import standardize_data_formats
dataset = load_dataset("Cyleux/gemma3n-conversational-reasoning", split="train[:3000]")
dataset = standardize_data_formats(dataset)
Schema:
conversations: ShareGPT-style list of turns with from and value
metadata columns are included for analysis and filtering
Notes:… See the full description on the dataset page: https://huggingface.co/datasets/Cyleux/gemma3n-conversational-reasoning.Finch-Collection-Gemini-3-Flash
Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks
A mid-training "practice phase" that teaches small open-source LLMs how to evolve solutions.
👋 This is the Gemini-3-Flash teacher variant of the Finch Collection — evolutionary search trajectories from the paper Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks, but with Gemini-3-Flash as the teacher mutation… See the full description on the dataset page: https://huggingface.co/datasets/minnesotanlp/Finch-Collection-Gemini-3-Flash.forbidden-backrooms-gemma-4-31B-it
Forbidden Backrooms: Gemma-4 31B Self-Chat
Self-chat transcripts and per-message embeddings for two role-inverted instances of Gemma-4-31B-it, comparing the official instruct checkpoint against an abliterated fine-tune of the same checkpoint. Both variants use identical int4 quantization served via Ollama, so quantization noise is not a confound between them.
The methodology follows Anthropic's Claude Opus 4 system card section on the "spiritual bliss attractor state." Leave two… See the full description on the dataset page: https://huggingface.co/datasets/alliedtoasters/forbidden-backrooms-gemma-4-31B-it.pie-gem5-pairs
PIE gem5-timed code optimization (src,tgt pairs)
C++ program-optimization data derived from the PIE dataset
("Learning Performance-Improving Code Edits"),
re-timed end-to-end with gem5 (x86 Skylake, syscall-emulation mode) at
per-test-case granularity. One row per official (source, target) program pair.
This dataset is reward-agnostic: it ships the full per-test-case reference timings and
case manifests so a downstream RL / eval pipeline decides at runtime how many cases to use… See the full description on the dataset page: https://huggingface.co/datasets/stablegradients/pie-gem5-pairs.bird-train-gemini3-flash
Dataset Card for Think2SQL-SFT
This dataset is a distilled Supervised Fine-Tuning (SFT) dataset designed to improve the reasoning capabilities of models in Text-to-SQL tasks.
It contains high-quality reasoning traces and SQL queries generated by Gemini 3 Flash.
Paper: Think2SQL: Blueprinting Reward Density and Advantage Scaling for Effective Text-To-SQL Reasoning
Base Benchmark: BIRD-Train
Dataset Description
The dataset consists of 9,428 high-quality traces, of… See the full description on the dataset page: https://huggingface.co/datasets/anonymous-2321/bird-train-gemini3-flash.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.ocn-empty-negations-generations-main-gemma4-qwen35
OCN OSS Model Generations
This dataset contains open-source model generations for prompts designed to elicit or suppress contrastive-negation framing.
Columns
prompt metadata from the OCN prompt bank;
model_id: Hugging Face model id;
model_family: model family;
model_stage: base, instruct, or other;
decoding: decoding configuration name;
seed: generation seed;
response: generated answer;
created_at: notebook run timestamp.
experiment_id: experiment cohort… See the full description on the dataset page: https://huggingface.co/datasets/ritwikraha/ocn-empty-negations-generations-main-gemma4-qwen35.taubench-gemini-traces
taubench-gemini-traces
Complete HTTP-level agentic traces from running taubench_gemini benchmark tasks through an instrumented reverse proxy.
Each trace captures full request/response pairs including system prompts, user messages, assistant responses, tool calls and results, and token usage metadata.
Stats
Total sessions: 115
Multi-turn sessions (2+ LLM calls): 115
Total records: 5744
Total LLM requests: 2872
Format
Raw JSONL traces from the instrumented… See the full description on the dataset page: https://huggingface.co/datasets/sammshen/taubench-gemini-traces.med-synth-questions-gemma-3-27b-deepseek-v4-flash
Med Synth Questions (Gemma-3 + DeepSeek V4 Flash)
Synthetic reasoning traces and answers for medical questions from openmed-community/med-synth-questions-gemma-3-27b-it. Each record contains a medical question with SYNTH-style reasoning and a generated answer by DeepSeek V4 Flash.
Dataset Summary
29,148 records (2 dupes + 3,410 incomplete/truncated removed from 32,560 source)
29,148 reasoning turns (99.2% format compliance)
Average 1,591 chars per reasoning trace… See the full description on the dataset page: https://huggingface.co/datasets/mkurman/med-synth-questions-gemma-3-27b-deepseek-v4-flash.caliber-extension-gemma4-e2b-grpo-rollouts
CALIBER Extension — Gemma4-E2B GRPO Rollouts
Training rollouts from matched GRPO arms on google/gemma-4-E2B-it
(new-prompt template, non-thinking, full bf16, max completion 1500, 150 steps).
Subsets
subset
arm
τ
prior
rows
mean reward_total
accuracy
full schema
caliber
vanilla CALIBER
0.0
—
1600
2.298
0.514
0.664
mink
Min-K% prior
1.0
mink_0.2
4800
2.506
0.520
0.680
minkpp
Min-K++% prior
1.0
minkpp_0.2
4800
2.637
0.541
0.726
Load:
from datasets… See the full description on the dataset page: https://huggingface.co/datasets/dmnsh/caliber-extension-gemma4-e2b-grpo-rollouts.gpt4o-coding-eval-by-gemini1_5flash-koTranslated llama-duo/gpt4o-coding-eval-by-gemini1_5flash using nayohan/llama3-instrucTrans-enko-8b.
This dataset is a raw translated dataset and contains repetitive sentences generated by the model, so it needs to be filtered.
pie-gem5-bysrc
PIE gem5-timed code optimization (per source program)
C++ program-optimization data derived from the PIE dataset
("Learning Performance-Improving Code Edits"),
re-timed end-to-end with gem5 (x86 Skylake, syscall-emulation mode) at
per-test-case granularity. One row per unique (problem, source) program; the best surviving target is kept as an oracle ceiling.
This dataset is reward-agnostic: it ships the full per-test-case reference timings and
case manifests so a downstream RL /… See the full description on the dataset page: https://huggingface.co/datasets/stablegradients/pie-gem5-bysrc.gemma-4-31b-it_writingbench-en100
google/gemma-4-31b-it — writingbench-en100
Model outputs from the micro-creativity inference suite.
Model: google/gemma-4-31b-it
Dataset: writingbench-en100 (100 items)
Part of collection: ZachW/llm-creativity-benchmarks
Generation config
temperature: 0.0
max_tokens: 8192
seed: 42
backend: vllm
Columns
Column
Description
task_id
Unique task identifier
input
The exact prompt sent to the model (after meta-prompt application)… See the full description on the dataset page: https://huggingface.co/datasets/ZachW/gemma-4-31b-it_writingbench-en100.nla-gemma4e2b-relabel-v1-corpus
Gemma-4-E2B layer-23 activation corpus, relabeled (v1)
1356 training rows for an activation verbalizer. Each row pairs a residual-stream
activation captured at layer 23 of google/gemma-4-E2B with a natural-language label
describing what the model must have integrated at that position to predict its next
token. This is the training set behind
Solshine/gemma-4-e2b-nla-L23-av-priordev-relabel-v1-wd3.
Why it exists
An audit of the previous version of this corpus found… See the full description on the dataset page: https://huggingface.co/datasets/Solshine/nla-gemma4e2b-relabel-v1-corpus.gemma4-qwen35-gsm8k-rollouts
Gemma 4 and Qwen3.5 GSM8K Rollouts
This dataset contains 3,957 saved generations from three complete runs over
the 1,319-example openai/gsm8k main test split:
Model
Rows
Strict match
Flexible extract
google/gemma-4-26B-A4B
1,319
33.28%
39.95%
google/gemma-4-E4B
1,319
26.23%
30.86%
Qwen/Qwen3.5-35B-A3B
1,319
15.92%
23.12%
Every row includes the exact five-shot prompt, model generation, reference
answer, strict and flexible correctness flags, pinned… See the full description on the dataset page: https://huggingface.co/datasets/dureduck/gemma4-qwen35-gsm8k-rollouts.gemma3-reasoning-dropin-context
Gemma3 Reasoning Drop-in (Context Preserved)
Drop-in dataset with task/input/expected_output where input includes prior turns so follow-up replies remain coherent.
Use:
from datasets import load_dataset
dataset = load_dataset("Cyleux/gemma3-reasoning-dropin-context", split="train[:10000]")
Stats:
{
"input": "data/functiongemma_upload/train.jsonl",
"output": "data/gemma3_reasoning_dropin_context/train.jsonl",
"rows": 2225,
"samples_total": 1214,
"rows_written": 2225… See the full description on the dataset page: https://huggingface.co/datasets/Cyleux/gemma3-reasoning-dropin-context.gemma-3-27b-it_writingbench-en100
google/gemma-3-27b-it — writingbench-en100
Model outputs from the micro-creativity inference suite.
Model: google/gemma-3-27b-it
Dataset: writingbench-en100 (100 items)
Part of collection: ZachW/llm-creativity-benchmarks
Generation config
temperature: 0.0
max_tokens: 8192
seed: 42
backend: vllm
Columns
Column
Description
task_id
Unique task identifier
input
The exact prompt sent to the model (after meta-prompt application)… See the full description on the dataset page: https://huggingface.co/datasets/ZachW/gemma-3-27b-it_writingbench-en100.whylab-gemini-2-5-docker-validation
🛈 Anonymity Notice (2026-05-12): The associated manuscript is currently under peer review at a double-blind venue. Author identity and venue-specific identifiers have been withheld throughout this README, the BibTeX templates, and the CITATION.cff block. The dataset itself remains CC-BY-4.0 and is independently citable via its Zenodo DOI 10.5281/zenodo.20018468. The author byline will be restored after the review outcome is announced.
DOI
This dataset is citable via DataCite DOI… See the full description on the dataset page: https://huggingface.co/datasets/neogenesislab/whylab-gemini-2-5-docker-validation.gemma3n-conversational-reasoning-with-tools
Gemma3N Conversational Reasoning With Embedded Tool Traces
Prepared for Unsloth Gemma3/Gemma3N conversational notebooks that expect ShareGPT conversations.
Multi-turn conversations are preserved.
Reasoning blocks (<think>...</think>) are preserved.
Tool call traces are preserved by embedding them in assistant text as tags:
<tool_call ...>...</tool_call>
<tool_response ...>...</tool_response>
Use:
from datasets import load_dataset
from unsloth.chat_templates import… See the full description on the dataset page: https://huggingface.co/datasets/Cyleux/gemma3n-conversational-reasoning-with-tools.gemma-4-e2b-deception-behavior-completions
Gemma-4-E2B deception & behavior completions
Consolidated 910-row corpus of (scenario prompt + Gemma-4-E2B-generated completion) pairs from earlier mechanistic-interpretability experiments. Each row captures the prompt the model saw and the text it actually produced; for a subset, Claude-Haiku-4-5 judge verdicts and SAE-feature labels are included.
The corpus is meant to be used as activation-extraction input for downstream interpretability work — Natural Language Autoencoder (NLA)… See the full description on the dataset page: https://huggingface.co/datasets/Solshine/gemma-4-e2b-deception-behavior-completions.Magpie-Gemma2-Pro-200K-Filtered-koTranslate Magpie-Align/Magpie-Gemma2-Pro-200K-Filtered using nayohan/llama3-instrucTrans-enko-8b.
This is a raw translation dataset. It needs to be filtered for repetitions generated by the model.
@misc{xu2024magpie,
title={Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing},
author={Zhangchen Xu and Fengqing Jiang and Luyao Niu and Yuntian Deng and Radha Poovendran and Yejin Choi and Bill Yuchen Lin},
year={2024},
eprint={2406.08464}… See the full description on the dataset page: https://huggingface.co/datasets/nayohan/Magpie-Gemma2-Pro-200K-Filtered-ko.gemma-4-e2b-nla-av_sft-v0_1_x-gemini-persona-audit
Gemma-4-E2B NLA AV-SFT Training Corpus (v0.1.x, Gemini persona+audit)
The 4,734-row AV-SFT training corpus for the v0.1.x Gemma-4-E2B NLA — a 9-source-family diversified expansion over the v0.0.x OpenWebText-only corpus. Labels generated by Gemini CLI following the persona+audit pipeline (Dr. Marisol Chen labels, Dr. Riley Otsuka audits).
This is the in-progress v0.1.x labeled training set. AR-SFT companion is still being labeled (~16% complete as of this dataset publish). When the… See the full description on the dataset page: https://huggingface.co/datasets/Solshine/gemma-4-e2b-nla-av_sft-v0_1_x-gemini-persona-audit.gemma-4-31b-it_aime-all
google/gemma-4-31b-it — aime-all
Model outputs from the micro-creativity inference suite.
Model: google/gemma-4-31b-it
Dataset: aime-all (933 items)
Part of collection: ZachW/llm-creativity-benchmarks
Generation config
temperature: 0.0
max_tokens: 32768
seed: 42
backend: vllm
Columns
Column
Description
task_id
Unique task identifier
input
The exact prompt sent to the model (after meta-prompt application)
raw_output
Full… See the full description on the dataset page: https://huggingface.co/datasets/ZachW/gemma-4-31b-it_aime-all.gemma-4-e2b-nla-eval-smoke
Gemma-4-E2B NLA smoke-eval (20-row held-out set)
A 20-row held-out subset of OpenWebText activations extracted from google/gemma-4-E2B at layer 23. Used as the canonical eval set for smoke-testing the v0.0.1 Gemma-4-E2B NLA pair on a fresh environment.
This dataset is a subset of the held-out rl.parquet evaluation set used for the v0.0.1 round-trip eval (n=50 attempted, 42 evaluated after 8 empty-output exclusions, cos 0.438 ± 0.054). The 20-row subset preserves the activation… See the full description on the dataset page: https://huggingface.co/datasets/Solshine/gemma-4-e2b-nla-eval-smoke.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-4-e2b-nla-ar_sft-v0_0_x-haiku-persona-audit
Gemma-4-E2B NLA AR-SFT Training Corpus (v0.0.x, Claude Haiku persona+audit)
The 696-row AR-SFT training corpus used for the Option B Gemma-4-E2B NLA pair. Labels generated by Claude Haiku 4.5 following the persona+audit pipeline — Dr. Marisol Chen (synthetic mech-interp expert) labels first, Dr. Riley Otsuka (synthetic senior editor) audits the labels.
This is the matched companion to the v0.0.x AV labeled corpus. The pair completes the first open-source non-Anthropic-team NLA… See the full description on the dataset page: https://huggingface.co/datasets/Solshine/gemma-4-e2b-nla-ar_sft-v0_0_x-haiku-persona-audit.