--- language: - en thumbnail: "https://pbs.twimg.com/media/FThx_rEWAAEoujW?format=jpg&name=medium" tags: - t5 - contrastive learning - ranking - decoding - metric learning - pytorch - text generation - retrieval license: "apache-2.0" datasets: - Wikipedia - PG19 - C4 - relic - ChapterBreak - HellaSwag - ROCStories metrics: - MAUVE - human --- ## Main repository https://github.com/martiansideofthemoon/rankgen ## What is RankGen? RankGen is a suite of encoder models (100M-1.2B parameters) which map prefixes and generations from any pretrained English language model to a shared vector space. RankGen can be used to rerank multiple full-length samples from an LM, and it can also be incorporated as a scoring function into beam search to significantly improve generation quality (0.85 vs 0.77 MAUVE, 75% preference according to humans annotators who are English writers). RankGen can also be used like a dense retriever, and achieves state-of-the-art performance on [literary retrieval](https://relic.cs.umass.edu/leaderboard.html). ## Setup **Requirements** (`pip` will install these dependencies for you) Python 3.7+, `torch` (CUDA recommended), `transformers` **Installation** ``` python3.7 -m virtualenv rankgen-venv source rankgen-venv/bin/activate pip install rankgen ``` Get the data [here](https://drive.google.com/drive/folders/1DRG2ess7fK3apfB-6KoHb_azMuHbsIv4?usp=sharing) and place folder in root directory. Alternatively, use `gdown` as shown below, ``` gdown --folder https://drive.google.com/drive/folders/1DRG2ess7fK3apfB-6KoHb_azMuHbsIv4 ``` Run the test script to make sure the RankGen checkpoint has loaded correctly, ``` python -m rankgen.test_rankgen_encoder --model_path kalpeshk2011/rankgen-t5-base-all ### Expected output 0.0009239262409127233 0.0011521980725477804 ``` ## Using RankGen Loading RankGen is simple using the HuggingFace APIs (see Method-2 below), but we suggest using [`RankGenEncoder`](https://github.com/martiansideofthemoon/rankgen/blob/master/rankgen/rankgen_encoder.py), which is a small wrapper around the HuggingFace APIs for correctly preprocessing data and doing tokenization automatically. You can either download [our repository](https://github.com/martiansideofthemoon/rankgen) and install the API, or copy the implementation from [below](#rankgenencoder-implementation). #### [SUGGESTED] Method-1: Loading the model with RankGenEncoder ``` from rankgen import RankGenEncoder, RankGenGenerator rankgen_encoder = RankGenEncoder("kalpeshk2011/rankgen-t5-xl-pg19") # Encoding vectors prefix_vectors = rankgen_encoder.encode(["This is a prefix sentence."], vectors_type="prefix") suffix_vectors = rankgen_encoder.encode(["This is a suffix sentence."], vectors_type="suffix") # Generating text # use a HuggingFace compatible language model generator = RankGenGenerator(rankgen_encoder=rankgen_encoder, language_model="gpt2-medium") inputs = ["Whatever might be the nature of the tragedy it would be over with long before this, and those moving black spots away yonder to the west, that he had discerned from the bluff, were undoubtedly the departing raiders. There was nothing left for Keith to do except determine the fate of the unfortunates, and give their bodies decent burial. That any had escaped, or yet lived, was altogether unlikely, unless, perchance, women had been in the party, in which case they would have been borne away prisoners."] # Baseline nucleus sampling print(generator.generate_single(inputs, top_p=0.9)[0][0]) # Over-generate and re-rank print(generator.overgenerate_rerank(inputs, top_p=0.9, num_samples=10)[0][0]) # Beam search print(generator.beam_search(inputs, top_p=0.9, num_samples=10, beam_size=2)[0][0]) ``` #### Method-2: Loading the model with HuggingFace APIs ``` from transformers import T5Tokenizer, AutoModel tokenizer = T5Tokenizer.from_pretrained(f"google/t5-v1_1-xl") model = AutoModel.from_pretrained("kalpeshk2011/rankgen-t5-xl-pg19", trust_remote_code=True) ``` ### RankGenEncoder Implementation ``` import tqdm from transformers import T5Tokenizer, T5EncoderModel, AutoModel class RankGenEncoder(): def __init__(self, model_path, max_batch_size=32, model_size=None, cache_dir=None): assert model_path in ["kalpeshk2011/rankgen-t5-xl-all", "kalpeshk2011/rankgen-t5-xl-pg19", "kalpeshk2011/rankgen-t5-base-all", "kalpeshk2011/rankgen-t5-large-all"] self.max_batch_size = max_batch_size self.device = 'cuda' if torch.cuda.is_available() else 'cpu' if model_size is None: if "t5-large" in model_path or "t5_large" in model_path: self.model_size = "large" elif "t5-xl" in model_path or "t5_xl" in model_path: self.model_size = "xl" else: self.model_size = "base" else: self.model_size = model_size self.tokenizer = T5Tokenizer.from_pretrained(f"google/t5-v1_1-{self.model_size}", cache_dir=cache_dir) self.model = AutoModel.from_pretrained(model_path, trust_remote_code=True) self.model.to(self.device) self.model.eval() def encode(self, inputs, vectors_type="prefix", verbose=False, return_input_ids=False): tokenizer = self.tokenizer max_batch_size = self.max_batch_size if isinstance(inputs, str): inputs = [inputs] if vectors_type == 'prefix': inputs = ['pre ' + input for input in inputs] max_length = 512 else: inputs = ['suffi ' + input for input in inputs] max_length = 128 all_embeddings = [] all_input_ids = [] for i in tqdm.tqdm(range(0, len(inputs), max_batch_size), total=(len(inputs) // max_batch_size) + 1, disable=not verbose, desc=f"Encoding {vectors_type} inputs:"): tokenized_inputs = tokenizer(inputs[i:i + max_batch_size], return_tensors="pt", padding=True) for k, v in tokenized_inputs.items(): tokenized_inputs[k] = v[:, :max_length] tokenized_inputs = tokenized_inputs.to(self.device) with torch.inference_mode(): batch_embeddings = self.model(**tokenized_inputs) all_embeddings.append(batch_embeddings) if return_input_ids: all_input_ids.extend(tokenized_inputs.input_ids.cpu().tolist()) return { "embeddings": torch.cat(all_embeddings, dim=0), "input_ids": all_input_ids } ```