datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
circuitlens-gemma-2-2btranscoder-descriptions-and-evaluations
CircuitLens & WeightLens: Transcoder Descriptions and Evaluations
This dataset contains automatically generated descriptions and evaluation metrics for Gemma-2-2B transcoders, produced using CircuitLens and WeightLens methods.
Methods
CircuitLens: https://github.com/egolimblevskaia/CircuitLens
WeightLens: https://github.com/egolimblevskaia/WeightLens
Dataset Structure
The dataset is organized by layers (0, 4, 7, 10, 12, 15, 18, 21, 23, 25), with each layer… See the full description on the dataset page: https://huggingface.co/datasets/egolimblevskaia/circuitlens-gemma-2-2btranscoder-descriptions-and-evaluations.diffing-stats-gemma-2-2b-crosscoder-l13-mu4.1e-02-lr1e-04 Contains maximum activating examples for all the features of our crosscoder trained on gemma 2 2B layer 13 available here: https://huggingface.co/Butanium/gemma-2-2b-crosscoder-l13-mu4.1e-02-lr1e-04/blob/main/README.md
base_examples.pt contains all the maximum examples of the feature on a subset of validation test of fineweb
chat_examples.pt is the same but for lmsys chat data
chat_base_examples.pt is a merge of the two above files.
All files are of the type dict[int, list[tuple[float… See the full description on the dataset page: https://huggingface.co/datasets/science-of-finetuning/diffing-stats-gemma-2-2b-crosscoder-l13-mu4.1e-02-lr1e-04.gemma3n-slicing-configsThis repository contains configurations to slice Gemma 3n E4B, which is enabled thanks to it being a MatFormer.
The E4B model can be sliced into small models, trading off quality and latency/compute requirements.
We recommend exploring the [MatFormer Lab](TODO: add link) to getting started with slicing Gemma 3n E4B yourself.
For each configuration, we calculate the MMLU accuracy.
Although these are not the only configurations possible, they are optimal configurations
identified by calculating… See the full description on the dataset page: https://huggingface.co/datasets/google/gemma3n-slicing-configs.gemma-llm-prompt-recoverydiffing-stats-gemma-2-9b-it-L20-k100-lr1e-04-Crosscodergemma_2b_outputs
Gemma 2B Green LLM Experiment Outputs
This dataset repository contains experiment artifacts for Gemma 2B green-LLM runs, including LoRA adapter checkpoints, metrics, predictions, carbon logs, and figures.
Contents
checkpoints/: LoRA adapter checkpoints for CE baseline and joint-loss variants.
metrics/: training histories, SQuAD and MMLU summaries, prediction CSVs, calibration tables, and surrogate weights.
logs/: run histories and carbon summary JSON files.
carbon/:… See the full description on the dataset page: https://huggingface.co/datasets/PhotonTJ/gemma_2b_outputs.diffing-stats-gemma-2-9b-it-L20-mu1.0e-01-lr1e-04-local-shuffling-CrosscoderLoss10k_reports_gemma_v2
Dataset Card for Financial Document Analysis Dataset
Dataset Description
This dataset comprises structured conversational entries designed to facilitate the training and evaluation of models that analyze and summarize financial documents. Each entry includes a conversation ID, a specific step in the conversation, a system-generated prompt, a user question, and the corresponding model-generated response.
Fields Overview
conv_id: Unique identifier for each… See the full description on the dataset page: https://huggingface.co/datasets/yatharth97/10k_reports_gemma_v2.diffing-stats-gemma-2-9b-it-DPO-L20-k100-lr1e-04-dpo-simpo-Crosscodermax-activating-examples-gemma-2-2b-l13-ckissanediffing-stats-gemma-2-9b-L20-k100-lr1e-04-base-it-Crosscoderisa_gemma
Dataset Card for Company-Specific Financial Analysis Dataset
Dataset Description
This dataset contains structured conversation data formatted to train and evaluate natural language processing models on tasks related to financial analysis of specific companies. Each entry in the dataset consists of a conversation ID, steps within the conversation, system prompts, user questions, and the corresponding model responses.
Fields Overview
conv_id: Unique identifier… See the full description on the dataset page: https://huggingface.co/datasets/yatharth97/isa_gemma.diffing-stats-gemma-2-9b-L20-k100-lr1e-04-Crosscoderroman-nepali-gemma-finalgemma-7b-itdiffing-stats-gemma-2-2b-L13-k100-lr1e-04-local-shuffling-CCLossremoved_gemma_treesGemma-news-hindiAIME-2024-Gemma-3-4bgemma-3-1b-pt-blind-spots
Gemma-3-1b-pt Blind Spots Dataset
Dataset Description
This dataset documents blind spots (systematic errors) found in
google/gemma-3-1b-pt,
a 1-billion-parameter pretrained base model (not instruction-tuned)
released by Google in March 2025 as part of the Gemma 3 family.
Each row contains:
Column
Description
id
Unique probe index
category
Type of reasoning tested
prompt
The input fed to the model (text-completion style)
expected_output
The… See the full description on the dataset page: https://huggingface.co/datasets/Junaid687/gemma-3-1b-pt-blind-spots.gemma-r1-testdiffing-stats-gemma-2-9b-L20-k100-lr1e-04-base-dpo-Crosscoderdiffing-stats-gemma3_1B-kansas_abortion-L19-k100-lr1e-03-x32-local-shuffling-Crosscodergemma-3-1b-pt-blind-spots
Blind Spots of google/gemma-3-1b-pt
Model Tested
Model: google/gemma-3-1b-ptParameters: 1BType: Pre-trained base language model (not instruction-tuned)Tested by: Toka-Tarek | Biotechnology graduate & Pharmacogenetics Lab Specialist
How I Loaded the Model
Tested on Google Colab (free T4 GPU, 16GB VRAM).
Note: torch.float16 caused numerical instability (NaN/inf errors)
on the T4 GPU, so torch.float32 was used instead for stable generation.
from huggingface_hub… See the full description on the dataset page: https://huggingface.co/datasets/Toka-Tarek/gemma-3-1b-pt-blind-spots.gemma-all_atten_0.7_0.005_wiki_sentencesgemma_ours_beta_0.7_0.1Gemma-27Bdiffing-stats-gemma-2-2b-L13-k100-lr1e-04-local-shuffling-Decoupledgemmamaingemma4-yoruba-blindspotModel Tested: https://huggingface.co/google/gemma-4-E2B-it
I loaded the model by going to the model’s page on Hugging Face, clicking the “Use this model” button, and then selecting Google Colab, which already provided the setup to run the model. In the model’s description, it states that it is multilingual, with a training dataset that includes content in over 140 languages. This made me curious to test whether all languages were used equally during training, especially less widely used… See the full description on the dataset page: https://huggingface.co/datasets/nthomas123/gemma4-yoruba-blindspot.