---

📄 Layer by Layer: Uncovering Hidden Representations in Language Models

📌 Read the Paper on arXiv

📚 Citation:

@article{skean2025layer,
  title={Layer by {L}ayer: {U}ncovering hidden representations in language models},
  author={Skean, Oscar and Arefin, Md Rifat and Zhao, Dan and Patel, Niket and Naghiyev, Jalal and LeCun, Yann and Shwartz-Ziv, Ravid},
  journal={ICML},
  year={2025}
}

---

🛠 Installation

Create a new conda environment and install the dependencies:

pip install -r requirements.txt

If you encounter issues with representation-itl (repitl), try installing it manually:

pip3 install -e git+https://github.com/uk-cliplab/representation-itl.git#egg=representation-itl

---

🧪 MTEB Results

🔧 The MTEB Harness

The MTEB-Harness.py script handles interactions between models and the MTEB benchmark. Key arguments:

• --model_family: Type of model (e.g. 'Pythia').
• --model_size: Model size variant (e.g. '14m').
• --revision: Model version to use (default: 'main').
• --evaluation_layer: Which layer to evaluate (e.g. -1 for final).
• --base_results_path: Where results will be saved.
• --purpose: Task type: 'run_tasks', 'run_entropy_metrics', or 'download_datasets'.
• --raise_error: Raise and stop on errors (True/False).

---

▶️ Running MTEB Benchmarks

Single Model, Single Layer

To run benchmarks on the final layer (--evaluation_layer -1) of a model:

python3 -u MTEB-Harness.py \
    --model_family "Pythia" \
    --model_size "14m" \
    --revision "main" \
    --evaluation_layer -1 \
    --base_results_path "experiments/results" \
    --purpose run_tasks

---

Single Model, All Layers

To benchmark all layers in a model:

MODEL_NAME="Pythia"
SIZE="410m"
MAX_LAYER=50  # Set high to be safe
REVISION="main"
PURPOSE="run_tasks"
RESULTS_PATH="experiments/results"

for layer in $(seq 0 $MAX_LAYER); do
    python MTEB-Harness.py \
        --model_family $MODEL_NAME \
        --model_size $SIZE \
        --revision $REVISION \
        --base_results_path $RESULTS_PATH \
        --purpose $PURPOSE \
        --evaluation_layer $layer
done

---

Multiple Models, All Layers

MODEL_NAME="Pythia"
MODEL_SIZES=("14m" "70m" "410m")
MAX_LAYER=50
REVISION="main"
PURPOSE="run_tasks"
RESULTS_PATH="experiments/results"

for size in ${MODEL_SIZES[@]}; do
    for layer in $(seq 0 $MAX_LAYER); do
        python MTEB-Harness.py \
            --model_family $MODEL_NAME \
            --model_size $size \
            --revision $REVISION \
            --base_results_path $RESULTS_PATH \
            --purpose $PURPOSE \
            --evaluation_layer $layer
    done
done

More scripts, including SLURM examples, are available in the slurm_scripts/ folder.

---

📊 Calculating Representation Metrics

To compute metrics like prompt entropy across all layers:

MODEL_NAME="Pythia"
MODEL_SIZES=("14m" "70m" "160m" "410m")
REVISION="main"
PURPOSE="run_entropy_metrics"

for size in ${MODEL_SIZES[@]}; do
    echo "Evaluating $MODEL_NAME-$size"
    python MTEB-Harness.py \
        --model_family $MODEL_NAME \
        --model_size $size \
        --revision $REVISION \
        --base_results_path "experiments/results" \
        --purpose run_entropy_metrics
done