datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
gemma4-serving-bench-data
Gemma 4 12B (QAT-Q4_0) — Serving-Behavior Test Data
Test data, charts, and the running research log from an autonomous research
loop characterizing and tuning a Gemma 4 12B QAT-Q4_0 model served via
llama.cpp/llamafile on a single RTX 3080 Ti. Every ~30 min the loop
summarizes findings, proposes a goal, tests it end-to-end, documents success or
failure, and publishes here + to GitHub.
Model under test: gemma-4-12b-it-qat-q4_0.gguf (Google, June 2026), 128K
ctx, f16 KV, MTP… See the full description on the dataset page: https://huggingface.co/datasets/SEBK4C/gemma4-serving-bench-data.eagle3-hidden-states-gemma4fineweb-edu-gemma4-1024
FineWeb-Edu — pre-tokenized for fast LM pretraining (Gemma tokenizer, ArrayRecord/Grain)
Pre-tokenized FineWeb-Edu
(sample/100BT), packed into fixed-length sequences and stored as
ArrayRecord shards for zero-overhead
streaming with Grain. No on-the-fly tokenization
at train time — you read int32 tokens straight off disk.
Format
Tokenizer: google/gemma-4-12B-it (vocab size 262144). Documents are
separated by the EOS token id 1.
Packing: the token stream is… See the full description on the dataset page: https://huggingface.co/datasets/mlnomad/fineweb-edu-gemma4-1024.gemma-4-e2b-atlas
gemma4-interpretability
Gemma materials-science interpretability research archive
Research records supporting Reading and Steering Materials Science-Mechanism Representations in an Open-Weight Language Model, Markus J. Buehler. Release identifier: paper-revision-2026-09-06.
This archive supplies the original observations, supporting state arrays, exact prompts, protocols, intervention records, statistics, analysis source, and generated research figures. It includes the original 4B readout and geometry… See the full description on the dataset page: https://huggingface.co/datasets/lamm-mit/gemma4-interpretability.gemma-4-pretokenized-tracesgemma-4-31b-sae-features
Gemma-4-31B Sparse Autoencoder Features
3,000 interpreted and verified SAE features across all 60 layers of Google's Gemma-4-31B-IT model.
What's in this dataset?
For each of the 60 transformer layers in Gemma-4-31B, we trained a TopK-64 Sparse Autoencoder with 43,008 features (8x expansion from d_model=5376). We then selected the 50 most interesting features per layer using SIPIT (Sparse Input-Token Invertibility Probe) scores, interpreted them with two independent LLMs… See the full description on the dataset page: https://huggingface.co/datasets/Adam1010/gemma-4-31b-sae-features.financial-english-source-corpus-gemma4-e2b-1280gemma-4-e4b-webvid-4K
gemma-4-e4b-webvid-4K
This dataset contains the webvid_upgraded.json annotations and the videos referenced by that file.
Source: https://huggingface.co/datasets/OpenGVLab/VideoChat2-IT/tree/main/video/vqa/webvid_qa.
Files
webvid_upgraded.json: upgraded WebVid QA/action annotations.
videos/: MP4 files referenced by webvid_upgraded.json.
All video paths in webvid_upgraded.json are relative to the dataset root and point into videos/, for example… See the full description on the dataset page: https://huggingface.co/datasets/bear7011/gemma-4-e4b-webvid-4K.gemma-4-e4b-it-atlas
juiceb0xc0de/gemma-4-e4b-it-atlas
A brain atlas for google/gemma-4-E4B-it, the instruction-tuned E4B member of the Gemma 4 family. This is not a chat dataset or a benchmark. It is an internal-mechanics map built by running activations through a corpus of prompts and scoring what each layer, component, head, and feature direction is doing.
If you want to know what sliding-window and full-attention layers actually do differently inside one model, how KV cache sharing splits a… See the full description on the dataset page: https://huggingface.co/datasets/juiceb0xc0de/gemma-4-e4b-it-atlas.gemma4-e2b-base-topk128-tracesdolma3_mix-150B-1025-merged-gemma4
dolma3_mix-150B-1025-merged (Gemma4)
Tokenized copy of the Dolma3 150B mix (dolma3_mix-150B-1025-merged) using the Gemma4 tokenizer.
Format
Megatron-LM indexed binaries: paired *_text_document.bin and *_text_document.idx files (448 files total, ~616 GiB).
These are not Hugging Face datasets Arrow/Parquet shards. Load them with Megatron / Megatron-LM indexed dataset readers.
Source name
Local / blob name: dolma3_mix-150B-1025-merged-gemma4
gemma-4-31b-it-qat-q4_0-unquantized-distribution-fidelity-768x2048-v1
gemma-4-31B-it-qat-q4_0-unquantized quantization analysis
Mean KL divergence against on-disk size
Scored under the distribution-fidelity laws, version 15. Read LAWS.md first: these numbers are comparable only within this artifact's token suite, geometry, and runtime identity, and not against any number produced elsewhere.
Each candidate directory holds its one-pager (report.md), its raw report, its compliance receipt, and its Law 14 attribution where one was produced. reference/… See the full description on the dataset page: https://huggingface.co/datasets/phaedawg/gemma-4-31b-it-qat-q4_0-unquantized-distribution-fidelity-768x2048-v1.gemma4-e4b-rl100-hf-bf16-sdpa-topk128-overlay
Gemma 4 E4B RL100 top-k-128 target overlay
Precomputed off-policy distillation targets for the E4B-RL-step-100 to E2B experiment.
Source traces: JWei05/gemma4-e4b-rl100-topk128-traces at revision 2b6e49a0a456ee9d67b16a1dc61785562bee90c9
Direction: Gemma 4 E4B RL step 100 teacher to Gemma 4 E2B base student
Target engine: Hugging Face BF16 SDPA full forward
Width: top-k 128
Stored target token IDs: int32
Stored target log-probabilities: float16
Causal alignment: response token… See the full description on the dataset page: https://huggingface.co/datasets/JWei05/gemma4-e4b-rl100-hf-bf16-sdpa-topk128-overlay.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.gemma-4-26b-a4b-it-distribution-fidelity-768x2048-v1
gemma-4-26B-A4B-it quantization analysis
Mean KL divergence against on-disk size
Scored under the distribution-fidelity laws, version 15. Read LAWS.md first: these numbers are comparable only within this artifact's token suite, geometry, and runtime identity, and not against any number produced elsewhere.
Each candidate directory holds its one-pager (report.md), its raw report, its compliance receipt, and its Law 14 attribution where one was produced. reference/ carries the… See the full description on the dataset page: https://huggingface.co/datasets/phaedawg/gemma-4-26b-a4b-it-distribution-fidelity-768x2048-v1.gemma4-e4b-rl100-topk128-traces2026_08_20_refinement_math_chess_gemma3_12b_gemma4_31b_transition_feedback_tokgemma-4-e2b-it-SAE
gemma-4-e2b-it — 35-layer SAE atlas
Sparse autoencoders on every decoder layer of gemma-4-e2b-it. Trained from scratch in one rolling pipeline with an event-aware controller. 35 layers, 49,152 features per layer, no per-layer hand-tuning.
The base model is a stubborn one. 15 sliding-window layers, then BAM no KV cache, thick and getting thicker the deeper you go. This atlas was built the whole way through it anyway.
What this is
Three months of work. My first… See the full description on the dataset page: https://huggingface.co/datasets/juiceb0xc0de/gemma-4-e2b-it-SAE.gemma4-german-tutor-data
German Tutor — grammar correction, conversation & flashcard data
The training set, evaluation suites, source lexicons and eval results behind
kessenma/gemma4-e4b-german-tutor-4bit
— a Gemma 4 E4B fine-tune that runs fully on-device (MLX, 4-bit) as the tutor in a German
learning app.
The fine-tune lifted the core grammar suite from 72% → 85%, halved missed errors
(17% → 9%), and cut false corrections (34% → 22%). Everything needed to reproduce those
numbers is in this repo.… See the full description on the dataset page: https://huggingface.co/datasets/kessenma/gemma4-german-tutor-data.gemma4_multi_ntp_datagemma4-31b-layer-study
Gemma 4 31B IT — Layer Analysis Study
A comprehensive 3-phase empirical study of which transformer layers in Google's Gemma 4 31B IT model are critical, redundant, or actively harmful to model predictions.
This repository contains the complete raw measurements, the probe set, the analysis scripts, and the interactive web visualization of findings.
What's Inside
File
Description
layer_stats.json
Phase A — per-layer Block Influence (BI), residual norms… See the full description on the dataset page: https://huggingface.co/datasets/KikoCis/gemma4-31b-layer-study.gemma_4_31b_it_gender_secret_female_no_cot_training_rolloutsalphadiana-swe-mini-opencode-gemma4-20260725-v04
AlphaDiana SWE-Bench Verified Mini result
Run ID: 20260725-swe_bench_verified_mini-opencode-gemma-4-31b-it-v04-proxyphase-mt131072
Benchmark: SWE-Bench Verified Mini
Agent/harness: opencode via AlphaDiana Podman SWE harness
Model: google/gemma-4-31B-it
Slurm job: mt131072
Summary from local inspection:
50 task rows
49 valid_scored
1 runtime_error
0 provider_error
18 correct
finish reasons: none
valid-only accuracy: 0.3673
completed-row accuracy: 0.3600
total-denominator… See the full description on the dataset page: https://huggingface.co/datasets/n-pelleriti/alphadiana-swe-mini-opencode-gemma4-20260725-v04.gemma4-agent-sft
gemma4-agent-sft
A clean, deduplicated, mixture-balanced tool-calling agent SFT dataset for fine-tuning google/gemma-4-26B-A4B-it, normalized from three agentic sources (Agent-Ark/Toucan-1.5M, open-thoughts/AgentTrove, nvidia/Nemotron-SFT-Agentic-v2).
Format (text, not pre-tokenized)
default config — 132,909 examples. Columns:
id, source, source_subset, tool_names (list)
messages — JSON string: list of {role, content, tool_calls, tool_responses}
tools — JSON… See the full description on the dataset page: https://huggingface.co/datasets/voidful/gemma4-agent-sft.2026_08_05_refinement_5env_gemma3_12b_gemma4_31b_tokgemma-4-e4b-it-500-refalphadiana-swe-mini-zeroclaw-gemma4-20260724-r6m2
AlphaDiana SWE-Bench Verified Mini result
Run ID: 20260724-swe_bench_verified_mini-zeroclaw-gemma-4-31b-it-v02
Benchmark: SWE-Bench Verified Mini
Agent/harness: zeroclaw via AlphaDiana Podman SWE harness
Model: google/gemma-4-31B-it
Slurm job: v02
Summary from local inspection:
50 task rows
46 valid_scored
4 runtime_error
0 provider_error
22 correct
finish reasons: none
valid-only accuracy: 0.4783
completed-row accuracy: 0.4400
total-denominator pass@1/accuracy: 0.4400… See the full description on the dataset page: https://huggingface.co/datasets/n-pelleriti/alphadiana-swe-mini-zeroclaw-gemma4-20260724-r6m2.2026_08_12_refinement_math_chess_gemma3_12b_gemma4_31b_raw_student_tok2026_08_11_refinement_5env_gemma3_12b_gemma4_31b_raw_student_tok