medfalconv2-1a-40b-lora

MedFalcon v2.1a 40b LoRA - Step 4500

Model Description

This a model check point release at 4500 steps. For evaluation use only! Limitations:

• LoRA output will be more concise than the base model
• Due to the size, base knowledge may be overwritten from falcon-40b
• Due to the size, more hardware may be required to load falcon-40b when using this LoRA

Architecture

nmitchko/medfalconv2-1a-40b-lora' is a large language model LoRa specifically fine-tuned for medical domain tasks. It is based on Falcon-40b at 40 billion parameters.

The primary goal of this model is to improve question-answering and medical dialogue tasks. It was trained using LoRA, specifically QLora, to reduce memory footprint.

See Training Parameters for more info This Lora supports 4-bit and 8-bit modes.

Requirements

bitsandbytes>=0.39.0
peft
transformers

Steps to load this model:

1. Load base model using transformers
2. Apply LoRA using peft

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

model = "tiiuae/falcon-40b"
LoRA = "nmitchko/medfalconv2-1a-40b-lora"

# If you want 8 or 4 bit set the appropriate flags
load_8bit = True

tokenizer = AutoTokenizer.from_pretrained(model)

model = AutoModelForCausalLM.from_pretrained(model,
    load_in_8bit=load_8bit,
    torch_dtype=torch.float16,
    trust_remote_code=True,
)

model = PeftModel.from_pretrained(model, LoRA)

pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
    device_map="auto",
)

sequences = pipeline(
   "What does the drug ceftrioxone do?\nDoctor:",
    max_length=200,
    do_sample=True,
    top_k=40,
    num_return_sequences=1,
    eos_token_id=tokenizer.eos_token_id,
)

for seq in sequences:
    print(f"Result: {seq['generated_text']}")

Training Parameters

The model was trained for 4500 steps or 1 epoch on a custom, unreleased dataset named medconcat. medconcat contains only human generated content and weighs in at over 100MiB of raw text.

The below bash script initiated training in 4bit mode for a rather large LoRA:

CURRENTDATEONLY=`date +"%b %d %Y"`

sudo nvidia-smi -i 1 -pl 250

export CUDA_VISIBLE_DEVICES=0

nohup python qlora.py \
    --model_name_or_path models/tiiuae_falcon-40b \
    --output_dir ./loras/medfalcon2.1a-40b \
    --logging_steps 100 \
    --save_strategy steps \
    --data_seed 42 \
    --save_steps 200 \
    --save_total_limit 40 \
    --evaluation_strategy steps \
    --eval_dataset_size 1024 \
    --max_eval_samples 1000 \
    --per_device_eval_batch_size 1 \
    --max_new_tokens 32 \
    --dataloader_num_workers 3 \
    --group_by_length \
    --logging_strategy steps \
    --remove_unused_columns False \
    --do_train \
    --lora_r 128 \
    --lora_alpha 256 \
    --lora_modules all \
    --double_quant \
    --quant_type nf4 \
    --bf16 \
    --bits 4 \
    --warmup_ratio 0.03 \
    --lr_scheduler_type constant \
    --gradient_checkpointing \
    --dataset="training/datasets/medconcat/" \
    --dataset_format alpaca \
    --trust_remote_code=True \
    --source_max_len 16 \
    --target_max_len 512 \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 16 \
    --max_steps 4500 \
    --eval_steps 1000 \
    --learning_rate 0.0001 \
    --adam_beta2 0.999 \
    --max_grad_norm 0.3 \
    --lora_dropout 0.05 \
    --weight_decay 0.0 \
    --seed 0 > "${CURRENTDATEONLY}-finetune-medfalcon2.1a.log" &