Emergence of Context Characteristics Sensitivity in Large Language Models

Code for the paper "Emergence of Context Characteristics Sensitivity in Large Language Models".

Overview

This repository investigates how LLMs' sensitivity to context characteristics evolves across instruction fine-tuning (IFT) stages: supervised fine-tuning (SFT), direct preference optimization (DPO), and reinforcement learning with verifiable rewards (RLVR).

Key findings:

• SFT consistently instills sensitivity toward easy-to-understand contexts (high lexical similarity, low perplexity, shorter length) across model families and datasets.
• DPO's effect on sensitivity is driven by characteristic differences between chosen and rejected responses in the training data. Careful dataset curation is needed to neutralize SFT-induced biases.
• RLVR broadly preserves DPO's direction of change.

Repository Structure

.
├── data/
│   ├── evaluation_dataset/          # ConflictQA, Context-Reliance, DRUID eval sets
│   └── property_detection/
│       └── evaluation_dataset/      # Property-annotated eval sets (.tsv)
├── src/
│   ├── get_model_predictions/
│   │   ├── get_predictions.py       # Model inference entrypoint
│   │   └── prompts.py               # Prompt templates
│   ├── property_detection/
│   │   ├── get_properties.py        # Heuristic property extraction (Jaccard, Flesch, length, etc.)
│   │   ├── get_perplexity.py        # Perplexity calculation via HuggingFace
│   ├── prepare_data.ipynb           # Dataset preprocessing
│   └── get_plots.ipynb              # Figure generation
└── requirements.txt

Setup

pip install -r requirements.txt

Additional dependencies not in requirements.txt:

• spacy with en_core_web_trf — for named entity recognition (python -m spacy download en_core_web_trf)

Set your HuggingFace token in a .env file:

HF_TOKEN=your_token_here

Usage

1. Extract Context Characteristics

python src/property_detection/get_properties.py \
    --data_path data/evaluation_dataset/context-reliance.csv \
    --save_folder data/output/ \
    --properties "claim_evidence_jaccard_sim flesch_reading_ease_score evidence_length claim_length uncertain_rate_lexicon"

Available properties: claim_evidence_jaccard_sim, claim_entity_overlap, flesch_reading_ease_score, evidence_length, claim_length

To compute perplexity:

python src/property_detection/get_perplexity.py <model_name> <data_path> [revision]

2. Run Model Inference

python -m src.get_model_predictions.get_predictions \
    --data_file data/evaluation_dataset/context-reliance.csv \
    --save_folder data/output/ \
    --use_evidence yes \
    --model_name meta-llama/Llama-3.2-1B \
    --prompt_name qa

Use --prompt_name claim_verification for DRUID; use qa for ConflictQA and Context-Reliance. Use --instruct for instruction-tuned models. Use --revision to evaluate a specific checkpoint.

3. Analysis & Plots

Open src/prepare_data.ipynb first for data preprocessing and src/get_plots.ipynb to reproduce the paper's figures.

Training Configuration

Llama-3.2-1B was fine-tuned using the Open-Instruct library on 4×A100 GPUs:

Stage	Learning Rate	Batch Size	Grad. Accum.	Epochs	Max Seq. Len
SFT	3e-5	2	16	2	4096
DPO	2.5e-6	2	16	1	2048

Checkpoints are saved every 1,000 steps (SFT) and every 250 steps (DPO).