Falcon 7B LLM Fine Tune Model

Model description

This model is a fine-tuned version of the tiiuae/falcon-7b model using the QLoRa library and the PEFT library.

Intended uses & limitations

How to use

• The model and tokenizer are loaded using the from_pretrained methods.
• The padding token of the tokenizer is set to be the same as the end-of-sentence (EOS) token.
• The generation_config is used to set parameters for generating responses, such as the maximum number of new tokens to generate and the temperature for the softmax function.
• The prompt is defined, encoded using the tokenizer, and passed to the model.generate method to generate a response.
• The generated response is decoded using the tokenizer and printed.

# Import necessary classes and functions
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftConfig, PeftModel

# Specify the model
PEFT_MODEL = "hipnologo/falcon-7b-qlora-finetune-chatbot"

# Load the PEFT config
config = PeftConfig.from_pretrained(PEFT_MODEL)

# Load the base model and tokenizer
model = AutoModelForCausalLM.from_pretrained(
    config.based_model_name_or_path,
    return_dict=True,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

# Set the padding token to be the same as the EOS token
tokenizer.pad_token = tokenizer.eos_token

# Load the PEFT model
model = PeftModel.from_pretrained(model, PEFT_MODEL)

# Set the generation parameters
generation_config = model.generation_config
generation_config.max_new_tokens = 200
generation_config.temperature = 0.7
generation_config.top_p = 0.7
generation_config.num_return_sequences = 1
generation_config.pad_token_id = tokenizer.eos_token_id
generation_config.eos_token_id = tokenizer.eos_token_id

# Define the prompt
prompt = """
<human>: How can I create an account?
<assistant>:
""".strip()
print(prompt)

# Encode the prompt
encoding = tokenizer(prompt, return_tensors="pt").to(model.device)

# Generate a response
with torch.inference_mode():
  outputs = model.generate(
      input_ids=encoding.input_ids,
      attention_mask=encoding.attention_mask,
      generation_config=generation_config,
  )

# Print the generated response
print(tokenizer.decode(outputs[0],skip_special_tokens=True))

Training procedure

The model was fine-tuned on the Ecommerce-FAQ-Chatbot-Dataset using the bitsandbytes quantization config:

• load_in_8bit: False
• load_in_4bit: True
• llm_int8_threshold: 6.0
• llm_int8_skip_modules: None
• llm_int8_enable_fp32_cpu_offload: False
• llm_int8_has_fp16_weight: False
• bnb_4bit_quant_type: nf4
• bnb_4bit_use_double_quant: True
• bnb_4bit_compute_dtype: bfloat16

Framework versions

• PEFT 0.4.0.dev0

Evaluation results

The model was trained for 80 steps, with the training loss decreasing from 0.184 to nearly 0. The final training loss was 0.03094411873175886.

• Trainable params: 2359296
• All params: 3611104128
• Trainable%: 0.06533447711203746

License

This model is licensed under Apache 2.0. Please see the LICENSE for more information.