--- license: apache-2.0 datasets: - BAAI/Infinity-Instruct language: - en --- # Infinity Instruct

Beijing Academy of Artificial Intelligence (BAAI)
[Paper][Code][🤗] (would be released soon)

Infinity-Instruct-3M-0625-Llama3-70B is an opensource supervised instruction tuning model without reinforcement learning from human feedback (RLHF). This model is just finetuned on [Infinity-Instruct-3M and Infinity-Instruct-0625](https://huggingface.co/datasets/BAAI/Infinity-Instruct) and showing favorable results on AlpacaEval 2.0 compared to GPT4-0613. ## **News** - 🔥🔥🔥[2024/07/09] We release the model weights of [InfInstruct-Mistral-7B 0625](https://huggingface.co/BAAI/Infinity-Instruct-3M-0625-Mistral-7B), [InfInstruct-Qwen2-7B 0625](https://huggingface.co/BAAI/Infinity-Instruct-3M-0625-Qwen2-7B), [InfInstruct-Llama3-8B 0625](https://huggingface.co/BAAI/Infinity-Instruct-3M-0625-Llama3-8B), [InfInstruct-Llama3-70B 0625](https://huggingface.co/BAAI/Infinity-Instruct-3M-0625-Llama3-70B), and [InfInstruct-Yi-1.5-9B 0625](https://huggingface.co/BAAI/Infinity-Instruct-3M-0625-Yi-1.5-9B). - 🔥🔥🔥[2024/07/09] We release the chat dataset [Infinity-Instruct-0625](https://huggingface.co/datasets/BAAI/Infinity-Instruct), it is a upgraded version of the Infinity-Instruct-0613. - 🔥🔥🔥[2024/06/28] We release the model weight of [InfInstruct-Llama3-70B 0613](https://huggingface.co/BAAI/Infinity-Instruct-3M-0613-Llama3-70B). It shows favorable results on AlpacaEval 2.0 compared to GPT4-0613 without RLHF. - 🔥🔥🔥[2024/06/21] We release the model weight of [InfInstruct-Mistral-7B 0613](https://huggingface.co/BAAI/Infinity-Instruct-3M-0613-Mistral-7B). It shows favorable results on AlpacaEval 2.0 compared to Mixtral 8x7B v0.1, Gemini Pro, and GPT-3.5 without RLHF. - 🔥🔥🔥[2024/06/13] We share the intermediate result of our data construction process (corresponding to the [InfInstruct-3M](https://huggingface.co/datasets/BAAI/Infinity-Instruct) in the table below). Our ongoing efforts focus on risk assessment and data generation. The finalized version with 10 million instructions is scheduled for release in late June. ## **Training Details**

Infinity-Instruct-3M-0625-Llama3-70B is tuned on Million-level instruction dataset [Infinity-Instruct](https://huggingface.co/datasets/BAAI/Infinity-Instruct). First, we apply the foundational dataset Infinity-Instruct-3M to improve the foundational ability (math & code) of Llama3-70B, and get the foundational instruct model Infinity-Instruct-3M-Llama3-70B. Then we finetune the Infinity-Instruct-3M-Llama3-70B to get the stronger chat model Infinity-Instruct-3M-0625-Llama3-70B. Here is the training hyperparamers. ```bash epoch: 3 lr: 5e-6 min_lr: 0 lr_warmup_steps: 40 lr_decay_style: cosine weight_decay: 0.0 adam_beta1: 0.9 adam_beta2: 0.95 global_batch_size: 528 clip_grad: 1.0 ``` Thanks to [FlagScale](https://github.com/FlagOpen/FlagScale), we could concatenate multiple training samples to remove padding token and apply diverse acceleration techniques to the traning procudure. It effectively reduces our training costs. We will release our code in the near future! ## **Benchmark** | **Model** | **MT-Bench** | **AlpacaEval2.0** | |:-------------------------------:|:------------:|:-----------------:| | GPT 3.5 Turbo 0613 | 8.4 | 22.7 | | Mixtral 8x7B v0.1 | 8.3 | 23.7 | | Gemini Pro | -- | 24.4 | | GPT4-0613 | 9.2 | 30.2 | | Llama-3-70B-Instruct | 9.0 | 34.4 | | InfInstruct-3M-0613-Llama3-70B* | 8.7 | 31.5 | | InfInstruct-3M-0625-Llama3-70B* | 8.9 | **38.0** | *denote the model is finetuned without reinforcement learning from human feedback (RLHF). We evaluate Infinity-Instruct-3M-0625-Llama3-70B on the two most popular instructions following benchmarks. Mt-Bench is a set of challenging multi-turn questions including code, math and routine dialogue. AlpacaEval2.0 is based on AlpacaFarm evaluation set. Both of these two benchmarks use GPT-4 to judge the model answer. AlpacaEval2.0 displays a high agreement rate with human-annotated benchmark, Chatbot Arena. The result shows that InfInstruct-3M-0613-Llama3-70B achieved 31.2 in AlpacaEval2.0, which is higher than the 30.4 of GPT4-0613 and the 34.4 of the official Llama-3-70B-Instruct although it does not yet use RLHF. ## **How to use** Infinity-Instruct-3M-0625-Llama3-70B adopt the same chat template of [Llama3-70B-instruct](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct): ```bash <|begin_of_text|><|start_header_id|>user<|end_header_id|> How are you?<|eot_id|><|start_header_id|>assistant<|end_header_id|> Hi!<|eot_id|><|start_header_id|>user<|end_header_id|> How are you?<|eot_id|><|start_header_id|>assistant<|end_header_id|> ``` To apply this model and template in conversation scenarios, you can refer to the following code: ```python from transformers import AutoModelForCausalLM, AutoTokenizer, LogitsProcessorList import torch device = "cuda" # the device to load the model onto model = AutoModelForCausalLM.from_pretrained("BAAI/Infinity-Instruct-3M-0625-Llama3-70B", torch_dtype=torch.bfloat16, device_map="auto" ) tokenizer = AutoTokenizer.from_pretrained("BAAI/Infinity-Instruct-3M-0625-Llama3-70B") prompt = "Give me a short introduction to large language model." messages = [ {"role": "user", "content": prompt} ] text = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True ) model_inputs = tokenizer([text], return_tensors="pt").to(device) logits_processor = LogitsProcessorList( [ MinLengthLogitsProcessor(1, eos_token_id=tokenizer.eos_token_id), TemperatureLogitsWarper(0.7), ] ) generated_ids = model.generate( model_inputs.input_ids, logits_processor=logits_processor, max_new_tokens=512 ) generated_ids = [ output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids) ] response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] print(response) ``` ## **Disclaimer** The resources, including code, data, and model weights, associated with this project are restricted for academic research purposes only and cannot be used for commercial purposes. The content produced by any version of Infinity Instruct is influenced by uncontrollable variables such as randomness, and therefore, the accuracy of the output cannot be guaranteed by this project. This project does not accept any legal liability for the content of the model output, nor does it assume responsibility for any losses incurred due to the use of associated resources and output results. ## ## **Citation** Our paper, detailing the development and features of the **Infinity Instruct** dataset and finetuned models, will be released soon on arXiv. Stay tuned! ``` @article{InfinityInstruct2024, title={Infinity Instruct}, author={Beijing Academy of Artificial Intelligence (BAAI)}, journal={arXiv preprint arXiv:2406.XXXX}, year={2024} } ```