metadata

license: apache-2.0
datasets:
  - GAIR/lima
language:
  - en
pipeline_tag: text-generation

LIMSTRAL 🇲🍋

Mistral 7B fine-tuned on LIMA

This model is a fine-tuned version of mistralai/Mistral-7B-v0.1 on the LIMA dataset for instruction following downstream task.

Training procedure

The model was loaded on 8 bits and fine-tuned on the LIMA dataset using the LoRA PEFT technique with the huggingface/peft library and trl/sft for 2 epochs on 1 x A100 (40GB) GPU.

SFT Trainer params:

trainer = SFTTrainer(
    model=model,
    train_dataset=train_ds,
    eval_dataset=test_ds,
    peft_config=peft_config,
    dataset_text_field="text",
    max_seq_length=2048,
    tokenizer=tokenizer,
    args=training_arguments,
    packing=False
)

LoRA config:

config = LoraConfig(
        lora_alpha=16,
        lora_dropout=0.1,
        r=64,
        bias="none",
        task_type="CAUSAL_LM",
        target_modules = ['q_proj', 'k_proj', 'down_proj', 'v_proj', 'o_proj', 'gate_proj', 'up_proj']
    )

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0002
train_batch_size: 2
eval_batch_size: 8
seed: 66
gradient_accumulation_steps: 64
total_train_batch_size: 128
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: constant
lr_scheduler_warmup_ratio: 0.03
num_epochs: 2
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss
1.7917	0.72	5	1.7604
1.7743	1.44	10	1.7217

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

model_id = "mrm8488/limstral-7B-v0.1"
tokenizer = "mrm8488/limstral-7B-v0.1"

model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(model_id)

model.resize_token_embeddings(len(tokenizer))

gen = pipeline("text-generation", model=model, tokenizer=tokenizer, device=0)

instruction = "[INST] Write an email to say goodbye to me boss [\INST]"
res = gen(instruction, max_new_tokens=512, temperature=0.3, top_p=0.75, top_k=40, repetition_penalty=1.2)
print(res[0]['generated_text'])

Framework versions

Transformers 4.35.0.dev0
Pytorch 2.1.0+cu118
Datasets 2.14.6
Tokenizers 0.14.1