Project: Understanding the importance of punctuations in information propogation. Token level interventions and layer swapping experiments to evaluate the model reasoning ability. Experiments done on gpt2, gemma-2b, and deepseek-1.3b.

• Trained Models:   - models/: Includes zipped model files.
      - gpt2: finetuned gpt2 model can be unzipped and used directly.
      - deepseek, gemma: LoRA adapters. Merge with base models using scripts/run_merge.py.

• Datasets:     - Available in Datasets_intervention/

• Visualizations:     - All experiments generate visualizations automatically.

---

Setup

1. Clone the Repository

git clone https://github.com/anonymous/reasoning-interventions.git
cd reasoning-interventions

2. Install Requirements

pip install -r requirements.txt

Make sure you also install:

• nnsight
• transformers
• peft
• huggingface_hub

---

Merging Adapter with Base Model (Gemma / DeepSeek)

To replicate results with Gemma or DeepSeek models, use the run_merge.py script:

python run_merge.py \
    --zip_path "models/gemma_adapter.zip" \
    --base_model "google/gemma-2b" \
    --output_path "models/gemma_merged" \
    --hf_token "your_huggingface_token"

This script:

• Unzips the LoRA adapter.
• Loads the base model from HuggingFace.
• Applies the adapter and saves the full merged model.

---

Running Experiments

1. Token-Level Intervention

To run interchange intervention experiments:

cd scripts
python run_intervention.py \
    --model_path <model_path> \
    --dataset_dir <directory having dataset> \ 
    --intervention_type <type of intervention being done> \ 
    --model_name <GPT2/DeepSeek/Gemma>

---

2. Rule-Based Reasoning Experiments

To run rule intervention experiments:

cd scripts
python rule_intervention.py \
    --model_name <gpt2,gemma,deepseek> \
    --model_path <model_path> \ 
    --dataset_dir "datasets_intervention/<dataset_name>.json" \
    --rule_type <all_rule,if_then_rule>

---

3. Necessity & Sufficiency Ablation

Scripts available in ablation/:

• run_gpt2_single.py – for last-layer ablations
• run_incremental_ablation.py – for incremental layer removals
• run_necessity_sufficiency.py – test necessity and sufficiency of punctuation tokens

Example:

python run_gpt2_singl_layer_ablation.py \ 
    --model_path <model_path> \
    --output_path <results_storage_path> \
    --model_name <gpt2,deepseek,gemma>

---

4. Rule-Based Layer Swap

python run_layer_swap.py \
    --swap_type <if_then,all> \
    --model_path <model_path> \
    --model_name <model_name> \
    --dataset_dir <dataset_dir>

---

Visualization

All experiments generate CSVs and plots in visualisations/.

Datasets

All datasets are present in:

datasets_intervention/

• rule_taker_subject.csv – for subject intervention
• rule_taker_adjective.csv – for adjective intervention
• rule_taker_full_sentence.csv – for dot, full sentence, necessity and sufficiency experiments
• two_sent_check.csv - for two sentence swaps.
• rule_inference.csv - if_then rule analysis
• rule_inference.csv - All rule analysis
• layer_swap_all.csv - layer swap analysis for All rules.
• layer_swap_if_then.csv - layer swap analysis for if_then rules.

---