Instruction Tuning with and without Context: Behavioral Shifts and Downstream Impact

[EACL 2026] This is the official codebase for Instruction Tuning with and without Context: Behavioral Shifts and Downstream Impact. It includes the training and inference code for reproducing our main experiments.

---

🛠️ Requirements

Install dependencies using:

pip install -r requirements.txt

---

📦 Data

🔹 Data

We use a filtered version of the Self-RAG dataset for training.

---

🚀 Training

📝 Text-Only Domain

To train using LLaMA 3 8B as the base model:

torchrun --nnodes 1 --master_port=29100 --nproc_per_node 8 train.py \
  --enable_fsdp --low_cpu_fsdp \
  --training_argument configs/training_configs/llama3_train.json \
  --num_epochs 3 \
  --dataset llava_llama3_selfrag_multi_dataset \
  --dist_checkpoint_folder llama3_basemodel \
  --batch_size_training 128 \
  --micro_batch_size 32 \
  --loss_mask_context context \
  --model_use_peft

Argument descriptions:

• --training_argument: Select config with the base model name. Check files under configs/training_configs.

• --model_name: Base model to fine-tune.

• --token_name: Tokenizer name (Instruct version for chat template compatibility).

• --dataset: Training dataset (aligned with your base model). Check the list of datasets under configs/datasets_dpr.py.

• --dist_checkpoint_folder: Folder to save checkpoints.

• --loss_mask_context: Choose from:

  • no_context – standard instruction tuning
  • context – CINGS (ours)
  • no_mask – CINGS without context masking

• --model_use_peft: Use LoRA for parameter-efficient fine-tuning (remove to train all parameters).

We have included examples in run_llama3.sh

---

🖼️ Vision-Language Domain

After training the language model, use the official LLaVA repo for vision-language alignment.

Update the following scripts:

• scripts/pretrain.sh
• scripts/finetune.sh

Replace model_name_or_path with the checkpoint folder from the text-only training step (dist_checkpoint_folder).

---

🔍 Inference

📝 Text-Only Domain

CUDA_VISIBLE_DEVICES=0 accelerate launch inference.py \
  --training_argument {training_argument}.json \
  --dataset {dataset} \
  --dist_checkpoint_folder {dist_checkpoint_folder} \
  --val_batch_size 1 \
  --add_docs \
  --model_use_peft

Use the same arguments as training. Only --dataset should be updated to point to your evaluation dataset (see configs/datasets.py for available options).

---

🖼️ Vision-Language Domain

We follow the evaluation process from the official LLaVA repo. See the evaluation guide for details.

---

📖 Citation

If you use this work, please cite:

@misc{lee2026instructiontuningcontextbehavioral,
      title={Instruction Tuning with and without Context: Behavioral Shifts and Downstream Impact}, 
      author={Hyunji Lee and Seunghyun Yoon and Yunjae Won and Hanseok Oh and Geewook Kim and Trung Bui and Franck Dernoncourt and Elias Stengel-Eskin and Mohit Bansal and Minjoon Seo},
      year={2026},
      eprint={2506.15480},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.15480}, 
}

---

🙏 Acknowledgements

This repository builds on Meta’s LLaMA Recipes, and part of the codes are from Official LLaVA Repo. We are grateful to the community and all contributors.

Point of Contact

For any questions about the implementation or content of the paper, you could contact hyunjil at cs.unc.edu