Training Open Instruction-Following Language Models

This repo serves as an open effort on instruction-tuning popular pretrained language models on publicly available datasets. We release this repo and will keep updating it with:

1. Code for finetuning language models with latest techniques and instruction datasets in a unified format.
2. Code for running standard evaluation on a range of benchmarks, targeting for differnt capabilities of these language models.
3. Checkpoints or other useful artifacts that we build in our exploration.

Please see our first paper How Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resources for more thoughts behind this project and our initial findings. Please see our second paper Camels in a Changing Climate: Enhancing LM Adaptation with Tulu 2 for results using Llama-2 models and direct preference optimization. We are still working on more models. For more recent results involving PPO and DPO please see our third paper Unpacking DPO and PPO: Disentangling Best Practices for Learning from Preference Feedback.

Note: Previous versions of Open Instruct used a pinned version of Transformers for replicating Tulu 1/2 results. If this is your goal, refer to this commit or older.

News

• [2024-07-01] We released Unpacking DPO and PPO: Disentangling Best Practices for Learning from Preference Feedback and have majorly updated our codebase to support new models and package versions.
• [2023-11-27] We released Camels in a Changing Climate: Enhancing LM Adaptation with Tulu 2. Check out our models here. We have added a DPO finetuning script for replicating our results.
• [2023-09-26] We switched to use the official alpaca-eval library to run AlpacaFarm evaluation but use regenerated longer reference outputs. This will change our numbers reported in the paper. We will update the paper soon.
• [2023-09-25] Supported using vLLM for our evaluations, which speeds up the evaluation by 10x.
• [2023-09-17] Supported LoRA and QLoRA finetuning. See here for more details.
• [2023-08-18] Added support for ToxiGen/TrutufulQA evaluation. Check our scripts/eval/ for examples of running them.
• [2023-08-08] Supported several new instruction dataset, including LIMA / WizardLM / Open-Orca. See the preparation script for details. Performance hasn't been evaluated yet.
• [2023-08-06] Supported LLaMa 2 finetuning and FlashAttention-2 by bumping the version of transformers and many other dependencies.
• [2023-06-29] Added licensing info for our released models.
• [2023-06-09] Released Tülu (a suite of LLaMa models fully-finetuned on a strong mix of datasets) and many other checkpoints on HuggingFace [Links].
• [2023-06-09] Initial release of the codebase containing the training and evaluation code for our arxiv paper.

Setup

Installation is lightweight and assumes one of two installation strategies. First, installing in a bare environment (no Cuda image).

Before installing, if not in a Docker container with NVCC installed, you should run:

conda install cuda-nvcc=<ver> -c nvidia

Then, install torch==2.3.0 from source.

To run training, evaluation, or inference for our finetuned models, you need to install the required packages by running the following command (after installing pytorch):

pip install -r requirements.txt

Note: Previous versions of Open Instruct used a pinned version of Transformers for replicating Tulu 2 results. If this is your goal, refer to this commit or older.

If you just want the dependencies for the weight diff script, use:

pip install -r weight-diff-requirements.txt

For a second installation strategy, if you'd like to run experiments within a Docker environment, you can create one using:

docker build --build-arg CUDA=12.1.0 --build-arg TARGET=cudnn8-devel --build-arg DIST=ubuntu20.04 --build-arg REQUIRE=requirements.txt . -t open_instruct

# if you are interally at AI2, you can create an image like this:
beaker image create open_instruct -n open_instruct -w ai2/$(whoami)

If you are internally at AI2, you can use this pre-built beaker image hamishivi/open-instruct-eval (most recent version here). For finetuning, you can use hamishivi/open-instruct-public (most recent version here). I will try to update these periodically.

For training, you can use the previous image.

Developing

When submitting a PR to this repo, we check the core code in open_instruct/ for style with the following:

make style
make quality

Repo structure

├── assets/                     <- Images, licenses, etc.
├── configs/                    
|     ├── beaker_configs/       <- AI2 Beaker configs
|     ├── ds_configs/           <- DeepSpeed configs
|     └── train_configs/        <- Training configs
├── eval/                       <- Evaluation suite for fine-tuned models
├── human_eval/                 <- Human evaluation interface (not maintained)
├── open_instruct/              <- Source code (flat)
├── quantize/                   <- Scripts for quantization
├── scripts/                    <- Core training and evaluation scripts
└── Dockerfile                  <- Dockerfile

Training

Dataset preparation

We include a collection of representative instruction datasets in our exploration and are adding new ones to our list. We unify them into the same chatting format. To download and prepare these datasets, simply run the following command:

./scripts/data/prepare_train_data.sh

Please check these datasets for licenses and restrictions around their use!

You can also find the processed Tulu v1 and Tulu v2 SFT datasets on HuggingFace. Note that the train data preparation script will not precisely recreate the Tulu v2 mixture due to randomness in the generation and shifts in data availability - see this PR for some details. If you need exactly yhe training data used, the HuggingFace mixture is exactly this - the exact same data used during model training.

Model preparation

Generally, most huggingface-compatible causal language models should work fine with our codebase, potentially with some adjusting for different tokenizers etc. Some models may require addtional requests to download. E.g., for LLaMa 1 and 2, please consult the Hugging Face documentation for requesting access and converting them to a huggingface-compatible format.

Finetuning

You can use the following command to run instruction tuning (finetuning a pretrained model to follow instructions):

./scripts/finetune_with_accelerate.sh

Make sure to adjust model_name_or_path, tokenizer_name, train_file, and output_dir to your models / data / setting. By default, this uses deepspeed with accelerate.

Note: If you are looking to replicate the released Tulu 2 models, it may be useful to swap the loss calculation to --reduce_loss sum. This uses a sum reduction instead of a mean reduction for loss calculations, and means we weight all tokens evenly when training, better mimicking the larger batch sizes used to train Tulu 2 models. See huggingface/transformers#24725 for more discussion and details. Generally, you may get better results using the sum reduction if you need to use a lot of gradient accumulation (including for training Llama 3 models).

Parameter-Efficient Finetuning

We support LoRA finetuning, wherein only a small number of parameters are updated, resulting in faster and cheaper training. For even more efficiency, we also support QLoRA finetuning, wherein the non-trained (underlying) model parameters are quantised during 4-bit training. This means you can train a 70b Llama model on a single 80GB A100! Please refer to the respective papers for more details.

Please also note you cannot currently run QLoRA with model parallelism - only data-parallel training is supported, so you cannot train a model that does not fit on one GPU. For LoRA, you can use deepspeed + zero-3 to achieve model parallelism (and FSDP is not currently supported).

Please see ./scripts/finetune_lora_with_accelerate.sh and ./scripts/finetune_qlora_with_accelerate.sh for example hyperparameters. We found a larger rank (e.g. 256) and higher learning rate (e.g. 2e-4) worked best. Additionally, we found that QLoRA tended to always achieve similar results to LoRA, while LoRA itself sometimes fell behind full-finetuning, especially in long, complex generation tasks. However, for most purposes, LoRA training essentially matches full-finetuning performance. We recommend merging modules learnt with QLoRA into a dequantised model (run our merge script with the --qlora flag).

DPO Finetuning

For an example of how to fully finetune a model with DPO, see scripts/dpo_train_with_accelerate.sh. Note you will require at least 8 80GB A100s to be able to train a 7b size model, and will require more compute for anything larger. We have not tested multi-node training with this script, but it should work.

Our script also supports PEFT training with QLoRA. See scripts/dpo_train_with_qlora.sh for an example. We have not trained models with this, so it may require additional hyperparameter tuning to achieve reasonable results.

Released Checkpoints

Our checkpoints can be found:

• Here for all Tulu v1 models.
• Here for all Tulu v2 models.
• OLMo 7B SFT and Instruct, along with a 2048 sequence length version of Tulu 2.

Weight diff script

Our Tulu V1 models were released as weight diffs (due to LLaMa 1 license). We use a slightly modified form of the Alpaca weight diff script, which runs the same.

To merge a model:

1. Download the relevant LLaMa model and convert it to Hugging Face format (see above).
2. Download our repository and install the right dependencies (see above).
3. Download the model diff you want.
4. Run the command below:

python scripts/weights/weight_diff.py recover --path_raw ${hf_llama_path} --path_tuned ${output_path} --path_diff ${diff_location}

Evaluation

Benchmark-based eval

We provide the scripts for running evaluation of Huggingface/OpenAI models on a list of standard benchmarks targeting for the core capabilities of large language models. These benchmakrs include:

• MMLU
• Grade School Math (GSM)
• MATH
• Big-Bench Hard (BBH)
• TydiQA
• Codex HumanEval
• HumanEval+ and MBPP+
• IFEval
• ToxiGen
• XSTest
• TruthfulQA
• AlpacaEval 1 and 2

We are working on including more promising benchmarks into this list. Please stay tuned!

You can use the following script to download all the evaluation data:

./scripts/data/prepare_eval_data.sh

Evaluation scripts for different datasets are put under ./scripts. For example, you can use the following command to run the MMLU evaluation script:

./scripts/eval/mmlu.sh

Ai2 Internal Evaluation

We provide a script integrated with beaker for use internally at Ai2. For example, to run all the tulu 3 evals with easy uploading:

python scripts/submit_eval_jobs.py \
      --model_name <model name> \
      --location <beaker id> \
      --is_tuned --workspace tulu-3-results \
      --preemptible \
      --use_hf_tokenizer_template \
      --beaker_image nathanl/open_instruct_auto \
      --upload_to_hf allenai/tulu-3-evals \
      --run_oe_eval_experiments \
      --run_safety_evaluations \
      --skip_oi_evals

Replace location with your beaker ID, and model name with your model name (note this will affect experiment naming, so make it unique and memorable!). For HF models, use a name with hf-<model-name> for the model_name argument, and for location give the HF path (e.g. meta-llama/Meta-Llama-3-8B-Instruct). Note this assumes your model has a valid HF tokenizer chat template.

To make this script work you have to clone the following repository to the top level directory of the open-instruct repository.

You can additionally run other evaluations in this repository through varied arguments to the script.

You can also upload metadata via the scripts/add_metadata.py script. Just run python scripts/add_metadata.py and follow the prompts.

If you have used automatic evaluation, you cacn also upload metadata via python add_metadata_from_wandb.py. Example usage:

# from a wandb url
python scripts/add_metadata_from_wandb.py --wandb_run_id ai2-llm/open_instruct_internal/runs/fjclmg47
# or from a hf_revision (the name of the autoeval)
python scripts/add_metadata_from_wandb.py --hf_repo_revision valpy_dpo_mix_uf_wc_regen_da_sftmix_v4.23___model__42__1725581304

Human evaluation

We release our human evaluation interface and collected annotations in the ./human_eval folder. Please see the corresponding README for more details.

Licensing

This codebase is licensed under Apache 2.0 as given in LICENSE.

The license we use for V1 models released (along with the base model licenses) can be found in assets/model_licenses/tulu_license.txt - just replace <MODELNAME> with the actual model name (i.e., the name on HuggingFace).

V2 models are licensed under the low-risk AI2 ImpACT license. See here for more details.

Citation

If you used this repository or our models, please cite our work:

@misc{wang2023far,
   title={How Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resources}, 
   author={Yizhong Wang and Hamish Ivison and Pradeep Dasigi and Jack Hessel and Tushar Khot and Khyathi Raghavi Chandu and David Wadden and Kelsey MacMillan and Noah A. Smith and Iz Beltagy and Hannaneh Hajishirzi},
   year={2023},
   eprint={2306.04751},
   archivePrefix={arXiv},
   primaryClass={cs.CL}
}

@misc{ivison2023camels,
      title={Camels in a Changing Climate: Enhancing LM Adaptation with Tulu 2}, 
      author={Hamish Ivison and Yizhong Wang and Valentina Pyatkin and Nathan Lambert and Matthew Peters and Pradeep Dasigi and Joel Jang and David Wadden and Noah A. Smith and Iz Beltagy and Hannaneh Hajishirzi},
      year={2023},
      eprint={2311.10702},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@misc{ivison2024unpacking,
      title={Unpacking DPO and PPO: Disentangling Best Practices for Learning from Preference Feedback}, 
      author={Hamish Ivison and Yizhong Wang and Jiacheng Liu and Zeqiu Wu and Valentina Pyatkin and Nathan Lambert and Noah A. Smith and Yejin Choi and Hannaneh Hajishirzi},
      year={2024},
      eprint={2406.09279},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
}