added config

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.json filter=lfs diff=lfs merge=lfs -text

config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94e73e086a5ed14149cee99a1aa3e2563ec7ab536c1653ff332999afa3520694
+size 546

scripts/train_model.py

@@ -1,264 +0,0 @@
-import gc
-
-import torch
-from torch.optim import AdamW
-import bitsandbytes as bnb
-from datasets import load_dataset, Dataset
-
-from transformers import (
-    AutoConfig,
-    AutoTokenizer,
-    AutoModelForCausalLM,
-    TrainingArguments,
-    Trainer,
-    DataCollatorForLanguageModeling,
-)
-
-
-def _batch_iterator():
-    ## code
-    # dataset = load_dataset('bigcode/programming-languages-keywords', split='train')
-    
-    # for row in dataset:
-    #     for n in row['keywords']:
-    #         yield n
-    
-    # del dataset
-    # gc.collect()
-    # return
-
-    # code
-    dataset = (
-        load_dataset('bigcode/the-stack-smol-xs', lang, split='train', trust_remote_code=True)
-        for lang in [
-            'ada', 'agda', 'alloy', 'antlr', 'applescript', 'assembly', 'augeas', 'awk', 'batchfile', 'bison', 'bluespec', 'c',
-            'c++', 'c-sharp', 'clojure', 'cmake', 'coffeescript', 'common-lisp', 'css', 'cuda', 'dart', 'dockerfile', 'elixir',
-            'elm', 'emacs-lisp','erlang', 'f-sharp', 'fortran', 'glsl', 'go', 'groovy', 'haskell','html', 'idris', 'isabelle', 'java', 
-            'java-server-pages', 'javascript', 'julia', 'kotlin', 'lean', 'literate-agda', 'literate-coffeescript', 'literate-haskell',
-            'lua', 'makefile', 'maple', 'markdown', 'mathematica', 'matlab', 'ocaml', 'pascal', 'perl', 'php', 'powershell', 'prolog',
-            'protocol-buffer', 'python', 'r', 'racket', 'restructuredtext', 'rmarkdown', 'ruby', 'rust', 'sas', 'scala', 'scheme', 
-            'shell', 'smalltalk', 'solidity', 'sparql', 'sql', 'stan', 'standard-ml', 'stata', 'systemverilog', 'tcl', 'tcsh', 'tex', 
-            'thrift', 'typescript', 'verilog', 'vhdl', 'visual-basic', 'xslt', 'yacc', 'zig'
-        ]
-    )
-
-    for d in dataset:
-        for row in d:
-            yield row['content']
-
-    del dataset
-    gc.collect()
-
-    # text
-    dataset = load_dataset('nampdn-ai/tiny-textbooks', split='train')
-
-    for row in dataset:
-        yield row['text']
-
-    del dataset
-    gc.collect()
-
-    ## text
-    # dataset = (
-    #     load_dataset('wikimedia/wikisource', lang, split='train')
-    #     for lang in ['20231201.ar', '20231201.as', '20231201.az', '20231201.ban', '20231201.be', '20231201.bg', '20231201.bn', '20231201.br', '20231201.bs', '20231201.ca', '20231201.cs', '20231201.cy', '20231201.da', '20231201.de', '20231201.el', '20231201.en', '20231201.eo', '20231201.es', '20231201.et', '20231201.eu', '20231201.fa', '20231201.fi', '20231201.fo', '20231201.fr', '20231201.gl', '20231201.gu', '20231201.he', '20231201.hi', '20231201.hr', '20231201.hu', '20231201.hy', '20231201.id', '20231201.is', '20231201.it', '20231201.ja', '20231201.jv', '20231201.kn', '20231201.ko', '20231201.la', '20231201.li', '20231201.lij', '20231201.lt', '20231201.mk', '20231201.ml', '20231201.mr', '20231201.nap', '20231201.nl', '20231201.no', '20231201.or', '20231201.pa', '20231201.pl', '20231201.pms', '20231201.pt', '20231201.ro', '20231201.ru', '20231201.sa', '20231201.sah', '20231201.sk', '20231201.sl', '20231201.sr', '20231201.su', '20231201.sv', '20231201.ta', '20231201.te', '20231201.th', '20231201.tr', '20231201.uk', '20231201.vec', '20231201.vi', '20231201.wa', '20231201.yi', '20231201.zh', '20231201.zh-min-nan']
-    # )
-    #
-    # for d in dataset:
-    #     for row in d['text']:
-    #         yield row
-    #
-    # del dataset
-    # gc.collect()
-
-    # text
-    dataset = (
-        load_dataset('xu-song/cc100-samples', lang, split='train')
-        for lang in ['am', 'ar', 'as', 'az', 'be', 'bg', 'bn', 'bn_rom', 'br', 'bs', 'ca', 'cs', 'cy', 'da', 'de', 'el', 'en', 'eo', 'es', 'et', 'eu', 'fa', 'ff', 'fi', 'fr', 'fy', 'ga', 'gd', 'gl', 'gn', 'gu', 'ha', 'he', 'hi', 'hi_rom', 'hr', 'ht', 'hu', 'hy', 'id', 'ig', 'is', 'it', 'ja', 'jv', 'ka', 'kk', 'km', 'kn', 'ko', 'ku', 'ky', 'la', 'lg', 'li', 'ln', 'lo', 'lt', 'lv', 'mg', 'mk', 'ml', 'mn', 'mr', 'ms', 'my', 'my_zaw', 'ne', 'nl', 'no', 'ns', 'om', 'or', 'pa', 'pl', 'ps', 'pt', 'qu', 'rm', 'ro', 'ru', 'sa', 'si', 'sc', 'sd', 'sk', 'sl', 'so', 'sq', 'sr', 'ss', 'su', 'sv', 'sw', 'ta', 'ta_rom', 'te', 'te_rom', 'th', 'tl', 'tn', 'tr', 'ug', 'uk', 'ur', 'ur_rom', 'uz', 'vi', 'wo', 'xh', 'yi', 'yo', 'zh-Hans', 'zh-Hant', 'zu']
-    )
-
-    for d in dataset:
-        for row in d['text']:
-            yield row
-
-    del dataset
-    gc.collect()
-
-    ## text
-    # dataset = (
-    #     load_dataset('csebuetnlp/xlsum', lang, split='train')
-    #     for lang in ['amharic', 'arabic', 'azerbaijani', 'bengali', 'burmese', 'chinese_simplified', 'chinese_traditional', 'english', 'french', 'gujarati', 'hausa', 'hindi', 'igbo', 'indonesian', 'japanese', 'kirundi', 'korean', 'kyrgyz', 'marathi', 'nepali', 'oromo', 'pashto', 'persian', 'pidgin', 'portuguese', 'punjabi', 'russian', 'scottish_gaelic', 'serbian_cyrillic', 'serbian_latin', 'sinhala', 'somali', 'spanish', 'swahili', 'tamil', 'telugu', 'thai', 'tigrinya', 'turkish', 'ukrainian', 'urdu', 'uzbek', 'vietnamese', 'welsh', 'yoruba']
-    # )
-    #
-    # for d in dataset:
-    #     for row in d['text']:
-    #         yield row
-    #
-    # del dataset
-    # gc.collect()
-
-    ## text
-    # dataset = load_dataset('recursal/SuperWikiNEXT-32B', split='train')
-    #
-    # for row in dataset['text']:
-    #     yield row
-    #
-    # del dataset
-    # gc.collect()
-
-    # code
-    dataset = load_dataset('m-a-p/CodeFeedback-Filtered-Instruction', split='train')
-
-    for row in dataset:
-        yield row['query'] + '\n' + row['answer']
-
-    del dataset
-    gc.collect()
-
-    # code
-    dataset = load_dataset('nampdn-ai/tiny-codes', split='train')
-    
-    for row in dataset:
-        yield row['prompt'] + '\n' + row['response']
-    
-    del dataset
-    gc.collect()
-
-    # math
-    dataset = load_dataset('ajibawa-2023/Maths-College', split='train')
-    
-    for row in dataset:
-        yield row['instruction'] + '\n' + row['output']
-    
-    del dataset
-    gc.collect()
-
-    # math
-    dataset = load_dataset('microsoft/orca-math-word-problems-200k', split='train')
-
-    for row in dataset:
-        yield row['question'] + '\n' + row['answer']
-
-    del dataset
-    gc.collect()
-
-    # text
-    dataset = load_dataset('mlabonne/FineTome-100k', split='train')
-
-    for row in dataset['conversations']:
-        yield '\n'.join(n['value'] for n in row)
-
-    del dataset
-    gc.collect()
-
-    # instruction
-    dataset = load_dataset('arcee-ai/agent-data', split='train')
-    
-    for row in dataset['conversations']:
-        yield '\n'.join(n['value'] for n in row)
-    
-    del dataset
-    gc.collect()
-
-    # instruction
-    dataset = (
-        load_dataset('cognitivecomputations/SystemChat-2.0', data_files='SystemChat_filtered.jsonl', split='train'),
-        load_dataset('cognitivecomputations/SystemChat-2.0', data_files='SystemChat_multilingual.jsonl', split='train'),
-    )
-    
-    for d in dataset:
-        for row in d['messages']:
-            yield '\n'.join(n['content'] for n in row)
-    
-    del dataset
-    gc.collect()
-
-    # emoji
-    dataset = load_dataset('badrex/llm-emoji-dataset', split='train')
-    
-    for row in dataset:
-        yield f'{row["character"]}\n{row["unicode"]}\n{row["short description"]}\n{row["tags"]}\n{row["LLM description"]}'
-    
-    del dataset
-    gc.collect()
-
-
-def batch_iterator():
-    for text in _batch_iterator():
-        for i in range(0, len(text), 2048):
-            chunk = text[i:i + 2048]
-            tokenized = tokenize_function(chunk)
-            yield tokenized
-
-
-def tokenize_function(text):
-    outputs = tokenizer(text, truncation=True, padding='max_length', max_length=2048)
-    outputs['labels'] = outputs['input_ids'].copy()
-    return outputs
-
-
-tokenizer = AutoTokenizer.from_pretrained('../')
-print(tokenizer)
-
-config = AutoConfig.from_pretrained('mistralai/Mistral-7B-Instruct-v0.3')
-config.bos_token_id = tokenizer.bos_token_id
-config.eos_token_id = tokenizer.eos_token_id
-config.unk_token_id = tokenizer.unk_token_id
-config.pad_token_id = tokenizer.pad_token_id
-config.hidden_size = 512
-config.intermediate_size = 1792 # int(512 * 3.5)
-config.max_position_embeddings = 32768 # 32 * 1024
-config.num_attention_heads = 12
-config.num_hidden_layers = 10
-config.num_key_value_heads = 4
-config.rope_theta = 1_000_000.0
-config.sliding_window = 4096
-config.torch_dtype = torch.bfloat16
-config.use_cache = False
-print(config)
-
-model = AutoModelForCausalLM.from_config(config)
-print(model)
-
-dataset = Dataset.from_generator(batch_iterator)
-print(dataset)
-
-data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
-print(data_collator)
-
-optimizer = bnb.optim.AdamW8bit(
-    model.parameters(),
-    lr=1e-5,
-    betas=(0.9, 0.95),
-    weight_decay=0.1,
-)
-print(optimizer)
-
-training_args = TrainingArguments(
-    output_dir='./mistral-custom',
-    num_train_epochs=3,
-    per_device_train_batch_size=1,
-    gradient_accumulation_steps=8,
-    warmup_steps=500,
-    learning_rate=1e-5,
-    fp16=False,
-    bf16=True,
-    logging_dir='./logs',
-    logging_steps=10,
-    evaluation_strategy='no',
-    save_strategy='epoch',
-    torch_compile=True,
-)
-print(training_args)
-
-trainer = Trainer(
-    model=model,
-    args=training_args,
-    train_dataset=dataset,
-    data_collator=data_collator,
-    optimizers=(optimizer, None)
-)
-print(trainer)
-
-trainer.train()
-trainer.save_model('./mistral-custom-final')