MohammadKhosravi/cefr-gemma-hidden-states-combined

CEFR Gemma Hidden States Combined Dataset

Overview

This dataset contains high-fidelity 1D latent representations extracted from the frozen google/gemma-2-2b transformer architecture. It maps 13,500 English sentences across the 6 discrete Common European Framework of Reference for Languages (CEFR) proficiency bands ($A1 \rightarrow C2$).

This resource was engineered explicitly to support Phase 1 of a Plug and Play Language Model (PPLM) architecture, allowing researchers to train downstream lightweight linear steering classifiers without hosting heavy multi-billion-parameter language assets in active GPU memory across multiple training epochs.

Source Data Configuration

To maximize stylistic variance, eliminate high-dimensional feature starvation, and build robust linguistic classification boundaries, this corpus integrates and homogenizes four distinct open-source CEFR streams:

1. UniversalCEFR/readme_en: Clean pedagogical prose, textbook layouts, and structured reading comprehension text.
2. UniversalCEFR/cefr_sp_en: A massive library (~10k rows) of short, authentic, everyday conversational sentences providing raw syntactic volume.
3. UniversalCEFR/cefr_asag_en: Automated Short Answer Grading text featuring actual non-native student responses, capturing structural learner errors.
4. UniversalCEFR/elg_cefr_en: Formal European Language Grid documentation capturing highly advanced B2/C1 bureaucratic configurations.

Dataset Structure

The dataset contains a single train split consisting of two core features:

Column Name	Data Type	Description
hidden_vector	Sequence(float32)	A 1D vector of size 2,304 generated via attention-aware mean pooling on Layer 26 activations.
cefr_level	String	The target categorical proficiency label (A1, A2, B1, B2, C1, C2, including normalized + tags).

Class Distribution (Bell Curve Layout)

The combined corpus naturally follows a Gaussian distribution profile:

• A1: 2.5%
• A2: 15.3%
• B1: 30.3%
• B2: 32.1%
• C1: 16.2%
• C2: 2.7%

Note for downstream training: Due to the inherent data sparsity at the margins (A1 and C2), it is highly recommended to enforce Class-Weighted Cross-Entropy loss functions or strategic data oversampling during classifier training loops.

Feature Extraction Methodology

The arrays were computed using a deterministic parallel prefill processing configuration:

1. Model Baseline: google/gemma-2-2b processed text inputs using a squared parallel sequence batch size of 16 under strict torch.no_grad() constraints.
2. Attention-Aware Pooling: Latent representations ($H_t$) were intercepted right out of Layer 26. An explicit attention mask was applied element-wise to discard <pad> tensor artifacts, taking the mathematical mean strictly across valid token contexts to avoid feature corruption.
3. Sanitation: All floating-point metrics were scrubbed using torch.nan_to_num to anchor out-of-bounds $FP16$ inference overflows into safe, uniform parameters.
4. Randomization: The entire combined layout was fully scrambled via a global shuffle (seed=42) to guarantee homogeneous distribution for subsequent batch optimizers.

Usage & Intended Applications

This dataset is optimized for training a standalone, single-layer linear classification matrix to act as a PPLM steering head.

from datasets import load_dataset
import torch

# Download the representations
dataset = load_dataset("MohammadKhosravi/cefr-gemma-hidden-states-combined", split="train")

# Extract sample vector
sample_vector = torch.tensor(dataset[0]["hidden_vector"])
print("Vector size:", sample_vector.shape) # Output: torch.Size([2304])