Mistral 7B Customer Support

🚀 Mistral-7B-Customer-Support-v1

This model is a fine - tuned version of Mistral-7B-Instruct-v0.2, tailored for the Customer Support domain, enabling fast and accurate responses to customer inquiries.

🚀 Quick Start

This model serves as the first step in Bitext’s two - step approach to LLM fine - tuning for creating chatbots, virtual assistants, and copilots in the Customer Support domain. It offers customers prompt and precise answers regarding their banking needs.

✨ Features

• Domain - specific fine - tuning: Optimized for the Customer Support domain, it can handle various support transactions.
• Customization potential: Allows users to further customize the model with a small amount of their own data.
• Rich training data: Trained on a comprehensive dataset covering multiple customer service intents.

💻 Usage Examples

Basic Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("bitext-llm/Mistral-7B-Customer-Support-v1")
tokenizer = AutoTokenizer.from_pretrained("bitext-llm/Mistral-7B-Customer-Support-v1")

inputs = tokenizer("<s>[INST] I want to change to the standard account [/INST] ", return_tensors="pt")
outputs = model.generate(inputs['input_ids'], max_length=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

📚 Documentation

Intended Use

• Recommended applications: Designed for the first step in Bitext’s two - step LLM fine - tuning approach for creating chatbots, virtual assistants, and copilots in the Customer Support domain, providing fast and accurate answers about banking needs.
• Out - of - scope: Not for general conversational purposes, and should not be used for medical, legal, or safety - critical advice.

Model Architecture

The model employs the MistralForCausalLM architecture with a LlamaTokenizer. It maintains the configuration of the base Mistral model but has been fine - tuned to better understand and generate responses related to customer service.

Training Data

The model was fine - tuned using the [bitext/Bitext - customer - support - llm - chatbot - training - dataset](https://huggingface.co/datasets/bitext/Bitext - customer - support - llm - chatbot - training - dataset), which is designed for question and answer interactions in the customer service sector. This dataset includes instructions and responses across a variety of customer service topics, ensuring that the model can handle a wide range of inquiries related to this field. The dataset covers 27 intents assigned to 10 categories such as cancel_order, place_order, change_order, and check_invoice. Each intent has around 1000 examples, illustrating a training process aimed at understanding and generating accurate responses for customer service interactions.

Training Procedure

Hyperparameters

• Optimizer: AdamW with betas=(0.9, 0.999) and epsilon = 1e - 08
• Learning Rate: 0.0002 with a cosine learning rate scheduler
• Epochs: 1
• Batch Size: 8
• Gradient Accumulation Steps: 4
• Maximum Sequence Length: 1024 tokens

Environment

• Transformers Version: 4.40.0.dev0
• Framework: PyTorch 2.2.1+cu121
• Tokenizers: Tokenizers 0.15.0

Training Results

Dataset Description

Overview

The dataset used for fine - tuning can train Large Language Models for both Fine Tuning and Domain Adaptation. It includes:

• Use Case: Intent Detection
• Vertical: Customer Service
• 27 intents assigned to 10 categories
• 26,872 question/answer pairs, around 1000 per intent
• 30 entity/slot types
• 12 different types of language generation tags

Categories and Intents

The dataset covers the following categories and intents:

• ACCOUNT: create_account, delete_account, edit_account, switch_account
• CANCELLATION_FEE: check_cancellation_fee
• DELIVERY: delivery_options
• FEEDBACK: complaint, review
• INVOICE: check_invoice, get_invoice
• NEWSLETTER: newsletter_subscription
• ORDER: cancel_order, change_order, place_order
• PAYMENT: check_payment_methods, payment_issue
• REFUND: check_refund_policy, track_refund
• SHIPPING_ADDRESS: change_shipping_address, set_up_shipping_address

Entities

The dataset includes various entities such as:

• {{Order Number}}, {{Invoice Number}}, {{Online Order Interaction}}, {{Online Payment Interaction}}, {{Online Navigation Step}}, {{Online Customer Support Channel}}, {{Profile}}, {{Profile Type}}, {{Settings}}, {{Online Company Portal Info}}, {{Date}}, {{Date Range}}, {{Shipping Cut - off Time}}, {{Delivery City}}, {{Delivery Country}}, {{Salutation}}, {{Client First Name}}, {{Client Last Name}}, {{Customer Support Phone Number}}, {{Customer Support Email}}, {{Live Chat Support}}, {{Website URL}}, {{Upgrade Account}}, {{Account Type}}, {{Account Category}}, {{Account Change}}, {{Program}}, {{Refund Amount}}, {{Money Amount}}, {{Store Location}}

Language Generation Tags

The dataset contains tags for various linguistic phenomena:

• Lexical Variation: Morphological (M), Semantic (L)
• Syntactic Structure Variation: Basic (B), Interrogative (I), Coordinated (C), Negation (N)
• Language Register Variations: Politeness (P), Colloquial (Q), Offensive (W)
• Stylistic Variations: Keyword (K), Abbreviations (E), Errors and Typos (Z)
• Other Tags: Indirect Speech (D), Regional Variations (G), Respect Structures (R), Code Switching (Y)

Limitations and Bias

• The model is fine - tuned on a domain - specific dataset and may not perform well outside the scope of customer service.
• Users should be aware of potential biases in the training data, as the model's responses may inadvertently reflect these biases. The dataset aims to cover general customer service inquiries, but biases may exist for specific use cases.

Ethical Considerations

This model should be used responsibly, considering ethical implications of automated customer service. It is important to ensure that the model's advice complements human expertise and adheres to relevant customer service guidelines.

Acknowledgments

This model was developed by Bitext and trained on infrastructure provided by Bitext.

📄 License

This model, "Mistral - 7B - Customer - Support - v1," is licensed under the Apache License 2.0 by Bitext Innovations International, Inc. This open - source license allows for free use, modification, and distribution of the model but requires that proper credit be given to Bitext.

Key Points of the Apache 2.0 License

• Permissibility: Users are allowed to use, modify, and distribute this software freely.
• Attribution: You must provide proper credit to Bitext Innovations International, Inc. when using this model, in accordance with the original copyright notices and the license.
• Patent Grant: The license includes a grant of patent rights from the contributors of the model.
• No Warranty: The model is provided "as is" without warranties of any kind.

You may view the full license text at [Apache License 2.0](http://www.apache.org/licenses/LICENSE - 2.0).

This licensing ensures the model can be used widely and freely while respecting the intellectual contributions of Bitext. For more detailed information or specific legal questions about using this license, please refer to the official license documentation linked above.

[](https://github.com/OpenAccess - AI - Collective/axolotl)

base_model: mistralai/Mistral-7B-Instruct-v0.2
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: true

hub_model_id: malmarjeh/Mistral-7B-instruct-v0.2

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: bitext/Bitext-customer-support-llm-chatbot-training-dataset
    type:
      system_prompt: 'You are an expert in customer support.'
      field_instruction: instruction
      field_output: response
      format: '[INST] {instruction} [/INST]'
      no_input_format: '[INST] {instruction} [/INST]'

#datasets:
#  - path: json
#    type: alpaca_w_system.load_open_orca
#data_files: file.zip

dataset_prepared_path:

val_set_size: 0.05
output_dir: ./qlora-out

adapter: qlora
lora_model_dir:

sequence_len: 1024
sample_packing: true
pad_to_sequence_len: true

eval_sample_packing: False

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

wandb_project: axolotl
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 8
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

warmup_steps: 10
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  bos_token: '<s>'
  eos_token: '</s>'
  unk_token: '<unk>'