--- license: apache-2.0 language: - en datasets: - instruction-pretrain/ft-instruction-synthesizer-collection --- # Instruction Pre-Training: Language Models are Supervised Multitask Learners (EMNLP 2024) This repo contains the **context-based instruction synthesizer** in our paper [Instruction Pre-Training: Language Models are Supervised Multitask Learners](https://huggingface.co/papers/2406.14491). We explore supervised multitask pre-training by proposing ***Instruction Pre-Training***, a framework that scalably augments massive raw corpora with instruction-response pairs to pre-train language models. The instruction-response pairs are generated by an efficient instruction synthesizer built on open-source models. In our experiments, we synthesize 200M instruction-response pairs covering 40+ task categories to verify the effectiveness of *Instruction Pre-Training*. ***Instruction Pre-Training* outperforms *Vanilla Pre-training* in both general pre-training from scratch and domain-adaptive continual pre-training.** In pre-training from scratch, *Instruction Pre-Training* not only improves pre-trained base models but also benefits more from further instruction tuning. In continual pre-training, *Instruction Pre-Training* enables Llama3-8B to be comparable to or even outperform Llama3-70B.

### [2024/11/29] 🤗 Introduce the multimodal version of instruction synthesizer at [AdaMLLM](https://huggingface.co/papers/2411.19930), for synthesizing visual instruction tasks 🤗 **************************** **Updates** **************************** * 2024/11/30: Released the multimodal version of the instruction synthesizer: [Visual Instruction Synthesizer](https://huggingface.co/AdaptLLM/Adapt-MLLM-to-Domains) * 2024/9/20: Our paper has been accepted by EMNLP 2024 main conference🎉 * 2024/9/11: Updated [FAQ on continual pre-training from Llama3](https://huggingface.co/instruction-pretrain/instruction-synthesizer) * 2024/8/29: Updated [guidelines](https://huggingface.co/instruction-pretrain/medicine-Llama3-8B) on evaluating any 🤗Huggingface models on the domain-specific tasks * 2024/7/31: Updated pre-training suggestions in the `Advanced Usage` section of [instruction-synthesizer](https://huggingface.co/instruction-pretrain/instruction-synthesizer) * 2024/7/15: We scaled up the pre-trained tokens from 100B to 250B, with the number of synthesized instruction-response pairs reaching 500M. The performance trend on downstream tasks throughout the pre-training process:

* 2024/6/21: Released the [paper](https://huggingface.co/papers/2406.14491), [code](https://github.com/microsoft/LMOps), and [resources](https://huggingface.co/instruction-pretrain) ## Resources **🤗 We share our data and models with example usages, feel free to open any discussions at [this page](https://huggingface.co/papers/2406.14491)! 🤗** - Thanks to the demo [davanstrien/instruction-synthesizer](https://huggingface.co/spaces/davanstrien/instruction-synthesizer) for implementing our approach - Context-Based Instruction Synthesizer: [instruction-synthesizer](https://huggingface.co/instruction-pretrain/instruction-synthesizer) - Fine-Tuning Data for the Synthesizer: [ft-instruction-synthesizer-collection](https://huggingface.co/datasets/instruction-pretrain/ft-instruction-synthesizer-collection) - General Models Pre-Trained from Scratch (on 100B tokes): - [InstructLM-500M](https://huggingface.co/instruction-pretrain/InstructLM-500M) - [InstructLM-1.3B](https://huggingface.co/instruction-pretrain/InstructLM-1.3B) - Domain-Specific Models Pre-Trained from Llama3-8B: - [Finance-Llama3-8B](https://huggingface.co/instruction-pretrain/finance-Llama3-8B) - [Biomedicine-Llama3-8B](https://huggingface.co/instruction-pretrain/medicine-Llama3-8B) - General Instruction-Augmented Corpora: [general-instruction-augmented-corpora](https://huggingface.co/datasets/instruction-pretrain/general-instruction-augmented-corpora) - Domain-Specific Instruction-Augmented Corpora (no finance data to avoid ethical issues): [medicine-instruction-augmented-corpora](https://huggingface.co/datasets/instruction-pretrain/medicine-instruction-augmented-corpora) ## Synthesize Instruction-Response Pairs to Augment Any Raw Corpora We conduct multitask fine-tuning on a language model to develop an instruction synthesizer capable of generating instruction-response pairs from any raw text. The fine-tuning data are available at [ft-instruction-synthesizer-collection](https://huggingface.co/datasets/instruction-pretrain/ft-instruction-synthesizer-collection)

### 1. Basic Usage: Synthesize instruction-response pairs based on a given raw text **💗 Here is an amazing demo that implements our approach: [davanstrien/instruction-synthesizer](https://huggingface.co/spaces/davanstrien/instruction-synthesizer) 💗**
Click to expand ```python from transformers import AutoModelForCausalLM, AutoTokenizer model = AutoModelForCausalLM.from_pretrained("instruction-pretrain/instruction-synthesizer") tokenizer = AutoTokenizer.from_pretrained("instruction-pretrain/instruction-synthesizer") # Put your raw text here: context = '''Free Fishing Weekend in NYS Slated This weekend (June 28th-29th) New Yorkers may fish for free without a license in any of the state's 7,500 lakes and ponds or 50,000 miles of rivers and streams. In addition, there are a number of free events and fishing clinics taking place across the state to encourage New Yorkers to enjoy the great outdoors. For more information, visit''' def parse_pred(pred): """Extract the list of instruction-response pairs from the prediction""" QA_str_list = pred.split('') if not pred.endswith(''): QA_str_list = QA_str_list[:-1] QA_list = [] raw_questions = [] for QA_str in QA_str_list: try: assert len(QA_str.split('')) == 2, f'invalid QA string: {QA_str}' Q_str, A_str = QA_str.split('') Q_str, A_str = Q_str.strip(), A_str.strip() assert Q_str.startswith(''), f'invalid question string: {Q_str} in QA_str: {QA_str}' assert len(A_str) > 0, f'invalid answer string in QA_str: {QA_str}' Q_str = Q_str.replace('', '').strip() assert Q_str.lower() not in raw_questions, f'duplicate question: {Q_str}' QA_list.append({'Q': Q_str, 'A': A_str}) raw_questions.append(Q_str.lower()) except: pass return QA_list def get_instruction_response_pairs(context): '''Prompt the synthesizer to generate instruction-response pairs based on the given context''' prompt = f' {context} \n\n' inputs = tokenizer(prompt, add_special_tokens=False, return_tensors="pt").input_ids.to(model.device) outputs = model.generate(input_ids=inputs, max_new_tokens=400, do_sample=False)[0] pred_start = int(inputs.shape[-1]) pred = tokenizer.decode(outputs[pred_start:], skip_special_tokens=True) return parse_pred(pred) # Get the generated instruction-response paris instruction_response_pairs = get_instruction_response_pairs(context) # Print out the results print(f'# Context:\n{context}\n') for index, pair in enumerate(instruction_response_pairs): print(f'## Instruction {index + 1}:\n{pair["Q"]}\n## Response {index + 1}:\n{pair["A"]}\n') ```
### 2. Advanced Usage: Convert Raw Corpora into Instruction-Augmented Corpora at Scale We use vLLM to accelerate the synthesis process. On a single A100-80GB GPU, it takes about 1 day to synthesize instruction-response pairs for 1 billion tokens of raw corpora.
Click to expand 1). Set up dependencies: ```bash git clone https://github.com/microsoft/LMOps.git cd LMOps/instruction_pretrain ``` Install vLLM with pip or from [source](https://vllm.readthedocs.io/en/latest/getting_started/installation.html#build-from-source): ```bash pip install vllm ``` 2). Synthesize and Templify Few-shot Examples for Pre-Training A one-shot example consists of a piece of raw text followed by its instruction-response pairs. We conduct multi-round inferece to synthesize few-shot examples: the instruction-response pairs of different raw texts share the same pattern. Suppose there are N pieces of raw text in the corpora, and you would like to covert them into M-shot examples: ```python from vllm import LLM, SamplingParams from utils.read_compre import get_dataset, cook_pt_entries, run # Put your list of raw texts here raw_texts = [ "Genetically and medically susceptible workers.\nThe likelihood of an individual becoming ill from a hazardous material or condition is strongly influenced by both their genetic makeup and their underlying state of health. Although the past decade has seen great advances in understanding human variation in health and genetic polymorphisms and in the diagnosis and treatment of disease, much less progress has been made in effectively using this information to protect worker health. Scientific evidence for increased susceptibility often is weak and rarely satisfies legal thresholds for sufficient risk to warrant exclusion from a particular job. When public safety is a major concern, many legally mandated exclusions are not well justified. Medical opinions about fitness to work should be based upon a systematic and credible analysis of the condition, its relationship to ability and risk for a particular job, and knowledge of possible accommodations. Conclusions should reflect the limitations of scientific knowledge and guidance from antidiscrimination legislation.", "Exclusive Breastfeeding for Twin Babies and Its Influencing Factors: A Study in East Java, Indonesia.\nThis study aimed to identify the factors that influence the success of exclusive breastfeeding in twins. This cross-sectional study was conducted on 184 mothers who had twins aged 6-23 months in Malang Raya, East Java, Indonesia and used the consecutive sampling technique. The data was collected through distributing questionnaires containing questions related to knowledge about exclusive breastfeeding, breastfeeding self-efficacy, and the support of family and certified health workers. Multinomial regression statistical test results show that the most influential factor for the success of exclusive breastfeeding with twins was breastfeeding self-efficacy (OR 0.111; 95% CI 0.033-0.387). A high level of breastfeeding self-efficacy can increase a mother's confidence to be able to provide exclusive breastfeeding for twins. This study suggests that nurses can provide breastfeeding counselling to improve breastfeeding self-efficacy."] N = len(raw_texts) # Number of raw texts M = 2 # M-shot example max_model_len = 4096 # max squence len of the LM you intend to pre-train max_new_tokens = 400 # max number of tokens for the augmented instruction-response pairs # Create a sampling params object. sampling_params = SamplingParams(temperature=0, max_tokens=max_new_tokens) # Load the model and tokenizer llm = LLM(model="instruction-pretrain/instruction-synthesizer", max_model_len=max_model_len) # 1. multi-round inference to get the prediction prev_examples = [] BSZ = (N+M-1)//M for round in range(M): cur_raw_texts = raw_texts[round*BSZ: (round+1)*BSZ] # load data split = get_dataset(prev_examples=prev_examples, cur_raw_texts=cur_raw_texts, max_model_len=max_model_len, max_new_tokens=max_new_tokens) prev_examples = run(split, llm, sampling_params) # 2. templify the data for subsequent pre-training instruction_augmented_texts = [] for idx, entry in enumerate(prev_examples): texts = cook_pt_entries(read_collection=entry, random_seed=idx+12345) # change random seed for each entry for diveristy instruction_augmented_texts.extend(texts) # 3. print out the instruction_augmented_texts for idx, text in enumerate(instruction_augmented_texts): print(text) # Now you can use `instruction_augmented_texts` for pre-training! ```
**Pre-Training Suggestions:** Except for the pre-training data, *Instruction Pre-Training* keeps all other settings the same as *Vanilla Pre-Training*. Therefore, you can easily use any training framework, such as [OLMo](https://github.com/allenai/OLMo) (for pre-training from scratch) and [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) (for continual pre-training), to train on the templified instruction-augmented corpora. 1. For general pre-training from scratch, we recommend setting M = 2 and mixing the instruction-augmented corpora with unchanged raw corpora. 2. For domain-adaptive continual pre-training, we recommend setting M = 3 and mixing the instruction-augmented corpora with general instructions from [OpenOrca](https://huggingface.co/datasets/Open-Orca/OpenOrca) at a 1:1 ratio (counted by tokens). Each example from OpenOrca is formulated as "{question} {response}", with a white-space used to connect the question and response. Let's try our method in continual pre-training for a quick start---it works easily! Feel free to ask for any suggestions at [this page](https://huggingface.co/papers/2406.14491); we will reply ASAP🤗! ## FAQ on Continual Pre-Training from LLama3 **Q1: Do you use the official Llama3 instruction prompt for pre-training?** No, the provided Llama3 instruction prompt is designed for the [instruction-tuned model](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct), but our continual pre-training is conducted on the [pre-trained base model](https://huggingface.co/meta-llama/Meta-Llama-3-8B) where only BOS (`<|begin_of_text|>`) and EOS (`<|end_of_text|>`) tokens are required. **Q2: For the general instructions from OpenOrca, do you concatenate each instruction with its output using '\n'?** No, as mentioned in the pre-training suggestions, we use a simple whitespace to concatenate each question with its response for the general instruction data from OpenOrca. This is because OpenOrca's data is already templated with diverse natural languge templates (such as those with `\n`), so a whitespace is sufficient to formulate the data. Note that when using our templated instruction-augmented texts, you don't need to add any concatenations. **Q3: What about those system prompts in OpenOrca?** We simply discard the system prompts. **To put it all together, the text before tokenization looks like this:** ```python general_instruction_response_text = "<|begin_of_text|>{question} {response}<|end_of_text|>" instruction_augmented_text = "<|begin_of_text|>{instruction augmented text}<|end_of_text|>" ``` Then, for tokenization, you don't need to add BOS and EOS token ids. The tokenization code looks like this: ```python text_ids = tokenizer(text, add_special_tokens=False, **kwargs).input_ids ``` ## Citation If you find our work helpful, please cite us: [Instruction Pre-Training](https://huggingface.co/papers/2406.14491) (EMNLP 2024) ```bibtex @article{cheng2024instruction, title={Instruction Pre-Training: Language Models are Supervised Multitask Learners}, author={Cheng, Daixuan and Gu, Yuxian and Huang, Shaohan and Bi, Junyu and Huang, Minlie and Wei, Furu}, journal={arXiv preprint arXiv:2406.14491}, year={2024} } ``` [Adapt LLM to Domains](https://huggingface.co/papers/2309.09530) (ICLR 2024) ```bibtex @inproceedings{ cheng2024adapting, title={Adapting Large Language Models via Reading Comprehension}, author={Daixuan Cheng and Shaohan Huang and Furu Wei}, booktitle={The Twelfth International Conference on Learning Representations}, year={2024}, url={https://openreview.net/forum?id=y886UXPEZ0} } ```