Text2Text Generation
Transformers
PyTorch
Hindi
mt5
Inference Endpoints
nreimers commited on
Commit
503282c
1 Parent(s): 5f3546d
README.md ADDED
@@ -0,0 +1,82 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: hi
3
+ datasets:
4
+ - unicamp-dl/mmarco
5
+ widget:
6
+ - text: "पाइथन एक सामान्य कार्यों के लिए उपयुक्त, उच्च स्तरीय प्रोग्रामिंग भाषा (General Purpose and High Level Programming language), इन्टरैक्टिव, ऑब्जेक्ट ओरिएन्टेड, स्क्रिप्टिंग भाषा है। इस भाषा को इस तरह से डिजाइन किया गया है ताकि इसमें लिखे गए कोड आसानी से पढ़े और समझे जा सकें।"
7
+
8
+ license: apache-2.0
9
+ ---
10
+
11
+ # doc2query/msmarco-hindi-mt5-base-v1
12
+
13
+ This is a [doc2query](https://arxiv.org/abs/1904.08375) model based on mT5 (also known as [docT5query](https://cs.uwaterloo.ca/~jimmylin/publications/Nogueira_Lin_2019_docTTTTTquery-v2.pdf)).
14
+
15
+ It can be used for:
16
+ - **Document expansion**: You generate for your paragraphs 20-40 queries and index the paragraphs and the generates queries in a standard BM25 index like Elasticsearch, OpenSearch, or Lucene. The generated queries help to close the lexical gap of lexical search, as the generate queries contain synonyms. Further, it re-weights words giving important words a higher weight even if they appear seldomn in a paragraph. In our [BEIR](https://arxiv.org/abs/2104.08663) paper we showed that BM25+docT5query is a powerful search engine. In the [BEIR repository](https://github.com/beir-cellar/beir) we have an example how to use docT5query with Pyserini.
17
+ - **Domain Specific Training Data Generation**: It can be used to generate training data to learn an embedding model. In our [GPL-Paper](https://arxiv.org/abs/2112.07577) / [GPL Example on SBERT.net](https://www.sbert.net/examples/domain_adaptation/README.html#gpl-generative-pseudo-labeling) we have an example how to use the model to generate (query, text) pairs for a given collection of unlabeled texts. These pairs can then be used to train powerful dense embedding models.
18
+
19
+ ## Usage
20
+ ```python
21
+ from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
22
+ import torch
23
+
24
+ model_name = 'doc2query/msmarco-hindi-mt5-base-v1'
25
+ tokenizer = AutoTokenizer.from_pretrained(model_name)
26
+ model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
27
+
28
+ text = "पाइथन एक सामान्य कार्यों के लिए उपयुक्त, उच्च स्तरीय प्रोग्रामिंग भाषा (General Purpose and High Level Programming language), इन्टरैक्टिव, ऑब्जेक्ट ओरिएन्टेड, स्क्रिप्टिंग भाषा है। इस भाषा को इस तरह से डिजाइन किया गया है ताकि इसमें लिखे गए कोड आसानी से पढ़े और समझे जा सकें।"
29
+
30
+
31
+ def create_queries(para):
32
+ input_ids = tokenizer.encode(para, return_tensors='pt')
33
+ with torch.no_grad():
34
+ # Here we use top_k / top_k random sampling. It generates more diverse queries, but of lower quality
35
+ sampling_outputs = model.generate(
36
+ input_ids=input_ids,
37
+ max_length=64,
38
+ do_sample=True,
39
+ top_p=0.95,
40
+ top_k=10,
41
+ num_return_sequences=5
42
+ )
43
+
44
+ # Here we use Beam-search. It generates better quality queries, but with less diversity
45
+ beam_outputs = model.generate(
46
+ input_ids=input_ids,
47
+ max_length=64,
48
+ num_beams=5,
49
+ no_repeat_ngram_size=2,
50
+ num_return_sequences=5,
51
+ early_stopping=True
52
+ )
53
+
54
+
55
+ print("Paragraph:")
56
+ print(para)
57
+
58
+ print("\nBeam Outputs:")
59
+ for i in range(len(beam_outputs)):
60
+ query = tokenizer.decode(beam_outputs[i], skip_special_tokens=True)
61
+ print(f'{i + 1}: {query}')
62
+
63
+ print("\nSampling Outputs:")
64
+ for i in range(len(sampling_outputs)):
65
+ query = tokenizer.decode(sampling_outputs[i], skip_special_tokens=True)
66
+ print(f'{i + 1}: {query}')
67
+
68
+ create_queries(text)
69
+
70
+ ```
71
+
72
+ **Note:** `model.generate()` is non-deterministic for top_k/top_n sampling. It produces different queries each time you run it.
73
+
74
+ ## Training
75
+ This model fine-tuned [google/mt5-base](https://huggingface.co/google/mt5-base) for 66k training steps (4 epochs on the 500k training pairs from MS MARCO). For the training script, see the `train_script.py` in this repository.
76
+
77
+ The input-text was truncated to 320 word pieces. Output text was generated up to 64 word pieces.
78
+
79
+ This model was trained on a (query, passage) from the [mMARCO dataset](https://github.com/unicamp-dl/mMARCO).
80
+
81
+
82
+
config.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "google/mt5-base",
3
+ "architectures": [
4
+ "MT5ForConditionalGeneration"
5
+ ],
6
+ "d_ff": 2048,
7
+ "d_kv": 64,
8
+ "d_model": 768,
9
+ "decoder_start_token_id": 0,
10
+ "dropout_rate": 0.1,
11
+ "eos_token_id": 1,
12
+ "feed_forward_proj": "gated-gelu",
13
+ "initializer_factor": 1.0,
14
+ "is_encoder_decoder": true,
15
+ "layer_norm_epsilon": 1e-06,
16
+ "model_type": "mt5",
17
+ "num_decoder_layers": 12,
18
+ "num_heads": 12,
19
+ "num_layers": 12,
20
+ "output_past": true,
21
+ "pad_token_id": 0,
22
+ "relative_attention_max_distance": 128,
23
+ "relative_attention_num_buckets": 32,
24
+ "tie_word_embeddings": false,
25
+ "tokenizer_class": "T5Tokenizer",
26
+ "torch_dtype": "float32",
27
+ "transformers_version": "4.18.0",
28
+ "use_cache": true,
29
+ "vocab_size": 250112
30
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:29e3ca84d4a028bd10a7c869e72308d1c2ccbb3c051aaa33ca5cb0e6b5c0b471
3
+ size 2329700301
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"eos_token": "</s>", "unk_token": "<unk>", "pad_token": "<pad>"}
spiece.model ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ef78f86560d809067d12bac6c09f19a462cb3af3f54d2b8acbba26e1433125d6
3
+ size 4309802
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4d3fca0dbb3a53bc1eddfc2e47ef441d7a94a70879e6750baddab04441a78305
3
+ size 16330621
tokenizer_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"eos_token": "</s>", "unk_token": "<unk>", "pad_token": "<pad>", "extra_ids": 0, "additional_special_tokens": null, "special_tokens_map_file": "/home/patrick/.cache/torch/transformers/685ac0ca8568ec593a48b61b0a3c272beee9bc194a3c7241d15dcadb5f875e53.f76030f3ec1b96a8199b2593390c610e76ca8028ef3d24680000619ffb646276", "name_or_path": "google/mt5-base", "sp_model_kwargs": {}, "tokenizer_class": "T5Tokenizer"}
train_script.py ADDED
@@ -0,0 +1,164 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import logging
3
+ from torch.utils.data import Dataset, IterableDataset
4
+ import gzip
5
+ import json
6
+ from transformers import Seq2SeqTrainer, AutoModelForSeq2SeqLM, AutoTokenizer, Seq2SeqTrainingArguments
7
+ import sys
8
+ from datetime import datetime
9
+ import torch
10
+ import random
11
+ from shutil import copyfile
12
+ import os
13
+ import wandb
14
+ import random
15
+ import re
16
+ from datasets import load_dataset
17
+ import tqdm
18
+
19
+
20
+ logging.basicConfig(
21
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
22
+ datefmt="%Y-%m-%d %H:%M:%S",
23
+ handlers=[logging.StreamHandler(sys.stdout)],
24
+ )
25
+
26
+ parser = argparse.ArgumentParser()
27
+ parser.add_argument("--lang", required=True)
28
+ parser.add_argument("--model_name", default="google/mt5-base")
29
+ parser.add_argument("--epochs", default=4, type=int)
30
+ parser.add_argument("--batch_size", default=32, type=int)
31
+ parser.add_argument("--max_source_length", default=320, type=int)
32
+ parser.add_argument("--max_target_length", default=64, type=int)
33
+ parser.add_argument("--eval_size", default=1000, type=int)
34
+ #parser.add_argument("--fp16", default=False, action='store_true')
35
+ args = parser.parse_args()
36
+
37
+ wandb.init(project="doc2query", name=f"{args.lang}-{args.model_name}")
38
+
39
+
40
+
41
+
42
+
43
+ def main():
44
+ ############ Load dataset
45
+ queries = {}
46
+ for row in tqdm.tqdm(load_dataset('unicamp-dl/mmarco', f'queries-{args.lang}')['train']):
47
+ queries[row['id']] = row['text']
48
+
49
+ """
50
+ collection = {}
51
+ for row in tqdm.tqdm(load_dataset('unicamp-dl/mmarco', f'collection-{args.lang}')['collection']):
52
+ collection[row['id']] = row['text']
53
+ """
54
+ collection = load_dataset('unicamp-dl/mmarco', f'collection-{args.lang}')['collection']
55
+
56
+ train_pairs = []
57
+ eval_pairs = []
58
+
59
+
60
+ with open('qrels.train.tsv') as fIn:
61
+ for line in fIn:
62
+ qid, _, did, _ = line.strip().split("\t")
63
+
64
+ qid = int(qid)
65
+ did = int(did)
66
+
67
+ assert did == collection[did]['id']
68
+ text = collection[did]['text']
69
+
70
+ pair = (queries[qid], text)
71
+ if len(eval_pairs) < args.eval_size:
72
+ eval_pairs.append(pair)
73
+ else:
74
+ train_pairs.append(pair)
75
+
76
+
77
+ print(f"Train pairs: {len(train_pairs)}")
78
+
79
+
80
+ ############ Model
81
+ model = AutoModelForSeq2SeqLM.from_pretrained(args.model_name)
82
+ tokenizer = AutoTokenizer.from_pretrained(args.model_name)
83
+
84
+ save_steps = 1000
85
+
86
+ output_dir = 'output/'+args.lang+'-'+args.model_name.replace("/", "-")+'-'+datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
87
+ print("Output dir:", output_dir)
88
+
89
+ # Write self to path
90
+ os.makedirs(output_dir, exist_ok=True)
91
+
92
+ train_script_path = os.path.join(output_dir, 'train_script.py')
93
+ copyfile(__file__, train_script_path)
94
+ with open(train_script_path, 'a') as fOut:
95
+ fOut.write("\n\n# Script was called via:\n#python " + " ".join(sys.argv))
96
+
97
+ ####
98
+
99
+ training_args = Seq2SeqTrainingArguments(
100
+ output_dir=output_dir,
101
+ bf16=True,
102
+ per_device_train_batch_size=args.batch_size,
103
+ evaluation_strategy="steps",
104
+ save_steps=save_steps,
105
+ logging_steps=100,
106
+ eval_steps=save_steps, #logging_steps,
107
+ warmup_steps=1000,
108
+ save_total_limit=1,
109
+ num_train_epochs=args.epochs,
110
+ report_to="wandb",
111
+ )
112
+
113
+ ############ Arguments
114
+
115
+ ############ Load datasets
116
+
117
+
118
+ print("Input:", train_pairs[0][1])
119
+ print("Target:", train_pairs[0][0])
120
+
121
+ print("Input:", eval_pairs[0][1])
122
+ print("Target:", eval_pairs[0][0])
123
+
124
+
125
+ def data_collator(examples):
126
+ targets = [row[0] for row in examples]
127
+ inputs = [row[1] for row in examples]
128
+ label_pad_token_id = -100
129
+
130
+ model_inputs = tokenizer(inputs, max_length=args.max_source_length, padding=True, truncation=True, return_tensors='pt', pad_to_multiple_of=8 if training_args.fp16 else None)
131
+
132
+ # Setup the tokenizer for targets
133
+ with tokenizer.as_target_tokenizer():
134
+ labels = tokenizer(targets, max_length=args.max_target_length, padding=True, truncation=True, pad_to_multiple_of=8 if training_args.fp16 else None)
135
+
136
+ # replace all tokenizer.pad_token_id in the labels by -100 to ignore padding in the loss.
137
+ labels["input_ids"] = [
138
+ [(l if l != tokenizer.pad_token_id else label_pad_token_id) for l in label] for label in labels["input_ids"]
139
+ ]
140
+
141
+
142
+ model_inputs["labels"] = torch.tensor(labels["input_ids"])
143
+ return model_inputs
144
+
145
+ ## Define the trainer
146
+ trainer = Seq2SeqTrainer(
147
+ model=model,
148
+ args=training_args,
149
+ train_dataset=train_pairs,
150
+ eval_dataset=eval_pairs,
151
+ tokenizer=tokenizer,
152
+ data_collator=data_collator
153
+ )
154
+
155
+ ### Save the model
156
+ train_result = trainer.train()
157
+ trainer.save_model()
158
+
159
+
160
+ if __name__ == "__main__":
161
+ main()
162
+
163
+ # Script was called via:
164
+ #python train_hf_trainer_multilingual.py --lang hindi
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:48252ad277be7f1af9bf7237a4a322f0592fa77e716b7c9575ba18b237b62880
3
+ size 3247