metadata

datasets:
  - OpenAssistant/oasst1
pipeline_tag: text-generation

Falcon-40b-chat-oasst1

Falcon-40b-chat-oasst1 is a chatbot-like model for dialogue generation. It was built by fine-tuning Falcon-40B on the OpenAssistant/oasst1 dataset. This model was fine-tuned in 4-bit using 🤗 peft adapters, transformers, and bitsandbytes.

The training relied on a recent method called "Low Rank Adapters" (LoRA), instead of fine-tuning the entire model you just have to fine-tune adapters and load them properly inside the model.
Training took approximately 10 hours and was executed on a workstation with a single NVIDIA A100-SXM 40GB GPU (via Google Colab).
See attached Notebook for the code (and hyperparams) used to train the model.

Model Summary

Model Type: Causal decoder-only
Language(s) (NLP): English (primarily)
Base Model: Falcon-40B (License: TII Falcon LLM License, commercial use ok-ed)
Dataset: OpenAssistant/oasst1 (License: Apache 2.0, commercial use ok-ed)

Model Date

May 30, 2023

Quick Start

To prompt the chat model, use the following format:

<human>: [Instruction]
<bot>:

Example Dialogue 1

Prompter:

"""<human>: My name is Daniel. Write a short email to my closest friends inviting them to come to my home on Friday for a dinner party, I will make the food but tell them to BYOB.
<bot>:"""

Falcon-40b-chat-oasst1:

[Coming]

Example Dialogue 2

Prompter:

<human>: Create a list of five things to do in San Francisco.\n
<bot>:

Falcon-40b-chat-oasst1:

[Coming]

Direct Use

This model has been finetuned on conversation trees from OpenAssistant/oasst1 and should only be used on data of a similar nature.

Out-of-Scope Use

Production use without adequate assessment of risks and mitigation; any use cases which may be considered irresponsible or harmful.

Bias, Risks, and Limitations

This model is mostly trained on English data, and will not generalize appropriately to other languages. Furthermore, as it is trained on a large-scale corpora representative of the web, it will carry the stereotypes and biases commonly encountered online.

Recommendations

We recommend users of this model to develop guardrails and to take appropriate precautions for any production use.

How to Get Started with the Model

Setup

# Install and import packages
!pip install -q -U bitsandbytes loralib einops
!pip install -q -U git+https://github.com/huggingface/transformers.git 
!pip install -q -U git+https://github.com/huggingface/peft.git
!pip install -q -U git+https://github.com/huggingface/accelerate.git

import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer

GPU Inference in 4-bit

This requires a GPU with at least 27GB memory.

# load the model
peft_model_id = "dfurman/falcon-40b-chat-oasst1"
config = PeftConfig.from_pretrained(peft_model_id)

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16
)

model = AutoModelForCausalLM.from_pretrained(
    config.base_model_name_or_path,
    return_dict=True,
    quantization_config=bnb_config,
    device_map={"":0},
    use_auth_token=True,
    trust_remote_code=True,
)

tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)
tokenizer.pad_token = tokenizer.eos_token

model = PeftModel.from_pretrained(model, peft_model_id)

# run the model
prompt = """<human>: My name is Daniel. Write a long email to my closest friends inviting them to come to my home on Friday for a dinner party, I will make the food but tell them to BYOB.
<bot>:"""

batch = tokenizer(
    prompt,
    padding=True,
    truncation=True,
    return_tensors='pt'
)
batch = batch.to('cuda:0')

with torch.cuda.amp.autocast():
    output_tokens = model.generate(
        input_ids = batch.input_ids, 
        max_new_tokens=200,
        temperature=0.7,
        top_p=0.7,
        num_return_sequences=1,
        pad_token_id=tokenizer.eos_token_id,
        eos_token_id=tokenizer.eos_token_id,
    )

# Inspect outputs
print('\n\n', tokenizer.decode(output_tokens[0], skip_special_tokens=True))

Reproducibility

See attached Notebook for the code (and hyperparams) used to train the model.

CUDA Info

CUDA Version: 12.0
GPU Name: NVIDIA A100-SXM
Max Memory: {0: "37GB"}
Device Map: {"": 0}

Package Versions Employed

torch: 2.0.1+cu118
transformers: 4.30.0.dev0
peft: 0.4.0.dev0
accelerate: 0.19.0
bitsandbytes: 0.39.0
einops: 0.6.1