selfrag/selfrag_llama2_7b

1---2license: mit3---4This model is a 7B [Self-RAG](https://selfrag.github.io/) model that generates outputs to diverse user queries as well as *reflection tokens* to call the retrieval system adaptively and criticize its own output and retrieved passages.  5 6Self-RAG is trained on our instruction-following corpora with interleaving passages and reflection tokens using the standard next-token prediction objective, enabling efficient and stable learning with fine-grained feedback.  7At inference, we leverage reflection tokens covering diverse aspects of generations to sample the best output aligning users' preferences. 8See full descriptions in See full descriptions in [our paper](https://arxiv.org/abs/2310.11511). 9 10## Usage11Here, we show an easy way to quickly download our model from HuggingFace and run with `vllm` with pre-given passages. Make sure to install dependencies listed at [self-rag/requirements.txt](https://github.com/AkariAsai/self-rag/blob/main/requirements.txt). 12To run our full inference pipeline with a retrieval system and fine-grained tree decoding, please use [our code](https://github.com/AkariAsai/self-rag). 13 14```py15from transformers import AutoTokenizer, AutoModelForCausalLM16from vllm import LLM, SamplingParams17 18model = LLM("selfrag/selfrag_llama2_7b", download_dir="/gscratch/h2lab/akari/model_cache", dtype="half")19sampling_params = SamplingParams(temperature=0.0, top_p=1.0, max_tokens=100, skip_special_tokens=False)20 21def format_prompt(input, paragraph=None):22  prompt = "### Instruction:\n{0}\n\n### Response:\n".format(input)23  if paragraph is not None:24    prompt += "[Retrieval]<paragraph>{0}</paragraph>".format(paragraph)25  return prompt26 27query_1 = "Leave odd one out: twitter, instagram, whatsapp."28query_2 = "Can you tell me the difference between llamas and alpacas?"29queries = [query_1, query_2]30 31preds = model.generate([format_prompt(query) for query in queries], sampling_params)32for pred in preds:33  print("Model prediction: {0}".format(pred.outputs[0].text))34# Model prediction: Twitter, Instagram, and WhatsApp are all social media platforms.[No Retrieval]WhatsApp is the odd one out because it is a messaging app, while Twitter and # Instagram are primarily used for sharing photos and videos.[Utility:5]</s> (this query doesn't require factual grounding; just skip retrieval and do normal instruction-following generation)35# Model prediction: Sure![Retrieval]<paragraph> ... (this query requires factual grounding, call a retriever)36 37# generate with retrieved passage38prompt = format_prompt("Can you tell me the difference between llamas and alpacas?", paragraph="The alpaca (Lama pacos) is a species of South American camelid mammal. It is similar to, and often confused with, the llama. Alpacas are considerably smaller than llamas, and unlike llamas, they were not bred to be working animals, but were bred specifically for their fiber.")39preds = model.generate([prompt], sampling_params)40print([pred.outputs[0].text for pred in preds])41# ['[Relevant]Alpacas are considerably smaller than llamas, and unlike llamas, they were not bred to be working animals, but were bred specifically for their fiber.[Fully supported][Utility:5]</s>']42```43 44## Input Format45As described in the `format_prompt` function, your input should be formed as 46```47### Instruction:\n{instruction}\n\n### Response:\n".format(instruction)48```49or50```51### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n"52```53If you have additional input. 54You can insert paragraphs anywhere after `### Response:\n"`, but make sure to mark paragraphs as paragraph tokens (i.e., `<paragraph>{0}</paragraph>`).55 56## Training details57Our training data is available at the HuggingFace dataset [selfrag_train_data](https://huggingface.co/datasets/selfrag/selfrag_train_data). 58See our official repository for the training details. 59We used 8 A100 40GB for training on the Stability HPC server.60 61## Citation and contact62If you use this model, please cite our work: 63```64@article{asai2023selfrag,65  author    = {Asai, Akari and Wu, Zeqiu and Wang, Yizhong and Sil, Avirup and Hajishirzi, Hannaneh},66  title     = {{Self-RAG}: Learning to Retrieve, Generate, and Critique through Self-Reflection},67  year      = {2023},68  journal   = { arXiv preprint arXiv:2310.11511 },69  URL       = {https://arxiv.org/abs/2310.11511}70}71```