README.md

---
license: apache-2.0
language:
- en
metrics:
- accuracy
library_name: transformers
pipeline_tag: text2text-generation
---

!pip -q install datasets
!pip install -q -U trl transformers accelerate git+https://github.com/huggingface/peft.git
!pip install -q bitsandbytes
!pip -q install  einops

import datasets
from datasets import load_dataset
dataset=load_dataset("diabolic6045/flanv2_cot_alpeca" , split="train")

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

model_name="TinyPixel/Llama-2-7B-bf16-sharded"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=bnb_config,
    trust_remote_code=True
)
model.config.use_cache = False

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token

filtered_dataset = []
for example in dataset:
    response = example['output']
    response_tokens = tokenizer.tokenize(response)
    if len(response_tokens) >= 100:
        filtered_dataset.append(example)

from datasets import Dataset
dict_of_lists = {key: [example[key] for example in filtered_dataset] for key in filtered_dataset[0]}
dataset = Dataset.from_dict(dict_of_lists)


from collections import defaultdict
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity


responses = [example["output"] for example in dataset]
tokenized_responses = [tokenizer.tokenize(response) for response in responses]

tfidf_vectorizer = TfidfVectorizer()
tfidf_matrix = tfidf_vectorizer.fit_transform([" ".join(tokens) for tokens in tokenized_responses])

cos_sim_matrix = cosine_similarity(tfidf_matrix, tfidf_matrix)

threshold = 0.95
grouped_responses = defaultdict(list)
for i in range(len(responses)):
    grouped_responses[i].append(i)  # Include the response itself in the group
    for j in range(i + 1, len(responses)):
        if cos_sim_matrix[i, j] > threshold:
            grouped_responses[i].append(j)
            grouped_responses[j].append(i)

deduplicated_responses = []
for group in grouped_responses.values():
    deduplicated_responses.append(responses[group[0]])

# Create a new dataset from deduplicated responses
deduplicated_dataset = Dataset.from_dict({"input": deduplicated_responses})

from peft import LoraConfig , get_peft_model

lora_alpha=16
lora_dropout = 0.1
lora_r= 64

peft_model=LoraConfig(
    lora_alpha= lora_alpha,
    lora_dropout = lora_dropout,
    r= lora_r,
    bias="none",
    task_type="CASUAL_LM"
)

from transformers import TrainingArguments

output_dir="./results"
per_device_train_batch_size = 4
gradient_accumulation_steps = 2
optim="paged_adamw_8bit"
save_steps = 100
logging_steps = 10
learning_rate = 2e-4
max_grad_norm = 1
max_steps = 100
warmup_ratio = 0.03
lr_scheduler_type = "constant"

training_argumet=TrainingArguments(
    output_dir=output_dir,
    per_device_train_batch_size=per_device_train_batch_size,
    gradient_accumulation_steps=gradient_accumulation_steps,
    optim=optim,
    save_steps=save_steps,
    logging_steps=logging_steps,
    learning_rate=learning_rate,
    fp16=True,
    max_grad_norm=max_grad_norm,
    max_steps=max_steps,
    warmup_ratio=warmup_ratio,
    group_by_length=True,
    lr_scheduler_type =lr_scheduler_type
)

from trl import SFTTrainer

max_seq_length= 512

model.train()
for name, param in model.named_parameters():
    if param.dtype in [torch.float32, torch.float64, torch.complex64, torch.complex128]:
        param.requires_grad = True

with torch.autograd.enable_grad():
  trainer = SFTTrainer(
      model=model,
      train_dataset=deduplicated_dataset,
      peft_config=peft_model,
      dataset_text_field="input",
      max_seq_length=max_seq_length,
      tokenizer=tokenizer,
      args=training_argumet
)

import argparse
import torch

parser = argparse.ArgumentParser(description='PyTorch Example')
parser.add_argument('--disable-cuda', action='store_true',
                    help='Disable CUDA')
args, unknown = parser.parse_known_args()
args.device = None
if not args.disable_cuda and torch.cuda.is_available():
    args.device = torch.device('cuda')
else:
    args.device = torch.device('cpu')


for name, module in trainer.model.named_modules():
    if "norm" in name:
        module = module.to(torch.float16)

trainer.train()

model_to_save=trainer.model.module if hasattr(trainer.model , 'module') else trainer.model
model_to_save.save_pretrained("output")

lora_config=LoraConfig.from_pretrained('output')
model=get_peft_model(model , lora_config)

dataset['input']

input="Given the sentence :A gathering of people with a young man playing a guitar. is it true that :A single woman is watching a band of guitar players.?"
device = "cuda:0"


inputs=tokenizer(input, return_tensors="pt").to(device)
output=model.generate(**inputs , max_new_tokens=50)
print(tokenizer.decode(output[0],skip_special_tokens=True))

from huggingface_hub import login
login()

model.push_to_hub("llama2_flan_v2_F.T")