monsoon-nlp's picture
tokenizer fix
4e7da13 verified
|
raw
history blame
3.33 kB
metadata
license: llama3
library_name: peft
language:
  - en
tags:
  - trl
  - sft
  - unsloth
  - generated_from_trainer
  - dna
base_model: gradientai/Llama-3-8B-Instruct-262k
model-index:
  - name: llama3-biotokenpretrain-kaniwa
    results: []

llama3-biotokenpretrain-kaniwa

This is a LoRA adapter.

The base model is the longer-context LLaMA-3-8b-Instruct developed by Gradient and Crusoe: gradientai/Llama-3-8B-Instruct-262k

The tokenizer has added "biotokens" ∎A, ∎C, ∎G, and ∎T.

The dataset was 0.5% of BYU's 2019 kaniwa (Chenopodium pallidicaule) genome, from https://genomevolution.org/coge/GenomeInfo.pl?gid=53872

The adapter was finetuned for 3 hours on an L4 GPU. The data was split into ~7k nucleotide snippets with an Alpaca like message format.

Training Notebook: https://colab.research.google.com/drive/1FKA3p_jnfRHYd-hqJdYmKn8MQpxec0t5?usp=sharing

Sample message:

Write information about the nucleotide sequence.

### Sequence:
∎G∎C∎C∎T∎A∎T∎A∎G∎T∎G∎T∎G∎T∎A∎G...

### Annotation:
Information about location in the kaniwa chromosome: >lcl|Cp5

Usage

Inference with DNA sequence

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

model = AutoPeftModelForCausalLM.from_pretrained("monsoon-nlp/llama3-biotokenpretrain-kaniwa", load_in_4bit=True).to("cuda")
tokenizer = AutoTokenizer.from_pretrained("monsoon-nlp/llama3-biotokenpretrain-kaniwa")
tokenizer.pad_token = tokenizer.eos_token # pad fix

qed = "∎" # from math symbols, used in pretraining
sequence = "".join([(qed + nt) for nt in "GCCTATAGTGTGTAGCTAATGAGCCTAGGTTATCGACCCTAATCT"])

inputs = tokenizer(f"{prefix}{sequence}{annotation}", return_tensors="pt")
outputs = model.generate(input_ids=inputs["input_ids"].to("cuda"), max_new_tokens=50)
sample = tokenizer.batch_decode(outputs, skip_special_tokens=False)[0]

LoRA finetuning on a new task

from transformers import AutoTokenizer
from trl import SFTTrainer
from unsloth import FastLanguageModel

model, _ = FastLanguageModel.from_pretrained(
    model_name = "monsoon-nlp/llama3-biotokenpretrain-kaniwa",
    max_seq_length = 7_000, # max 6,000 bp for AgroNT tasks
    dtype = None,
    load_in_4bit = True,
    resize_model_vocab=128260, # includes biotokens
)
tokenizer = AutoTokenizer.from_pretrained("monsoon-nlp/llama3-biotokenpretrain-kaniwa")
tokenizer.pad_token = tokenizer.eos_token # pad fix

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
...
)

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 0.0002
  • train_batch_size: 1
  • eval_batch_size: 8
  • seed: 3407
  • gradient_accumulation_steps: 4
  • total_train_batch_size: 4
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 5
  • training_steps: 280

Framework versions

  • PEFT 0.10.0
  • Transformers 4.40.2
  • Pytorch 2.2.1+cu121
  • Datasets 2.19.1
  • Tokenizers 0.19.1

Genome Citation

Mangelson H, et al. The genome of Chenopodium pallidicaule: an emerging Andean super grain. Appl. Plant Sci. 2019;7:e11300. doi: 10.1002/aps3.11300