Dynamic Tuning LLM

Dynamic information density tuning experiments for T5-based language models.

Quickstart

bash setup_env.sh
source .venv/bin/activate
python dynamic_tuning.py

Each run produces checkpoints in outputs/ and writes an eval_config.json that can be used with lm-eval:

lm_eval --config_file eval_config.json

Project Structure

• dynamic_tuning.py – training script with five dynamic nudging strategies.
• setup_env.sh – convenience script to create a virtual environment and install dependencies.
• outputs/ – checkpoints saved per strategy (created at runtime).
• eval_config.json – generated after training for quick evaluation with lm-eval.

Strategies

The script cycles through five strategies that adapt the model during training:

1. activation_entropy
2. embedding_separation
3. cross_layer_diversity
4. attention_variance
5. representation_sparsity

Each strategy corresponds to a distinct information-density heuristic applied to the T5 encoder.

Dataset

The experiments fine-tune t5-base on the COPA task from SuperGLUE. Prompts are formatted as a conditional generation problem where the model must output the correct choice.