Qra-1b-dolly-instruction-0.1

Qra-1b-dolly-instruction-0.1

This model if a fine-tuned version of OPI-PG/Qra-1b on the s3nh/alpaca-dolly-instruction-only-polish dataset.

Model Description

Trained from OPI-PG/Qra-1b

Intended uses & limitations

This model has been fine-tuned for question-answering task. It is possible to use it as a chat, but it doesn't work well because the dataset did not contain conversations.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline

model_id = "nie3e/Qra-1b-dolly-instruction-0.1"
device = "cuda" if torch.cuda.is_available() else "cpu"

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
)
tokenizer = AutoTokenizer.from_pretrained(model_id)
pipe = pipeline(
    "text-generation", model=model, tokenizer=tokenizer, device=device
)

def get_answer(system_prompt: str, user_prompt: str) -> str:
    input_msg = [
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": user_prompt}
    ]
    prompt = pipe.tokenizer.apply_chat_template(
        input_msg, tokenize=False,
        add_generation_prompt=True
    )
    outputs = pipe(
        prompt, max_new_tokens=512, do_sample=False, temperature=0.1, top_k=50,
        top_p=0.1, eos_token_id=pipe.tokenizer.eos_token_id,
        pad_token_id=pipe.tokenizer.pad_token_id
    )
    return outputs[0]['generated_text'][len(prompt):].strip()

print(
        get_answer(
        system_prompt="Jesteś przyjaznym chatbotem",
        user_prompt="Napisz czym jest dokument architectural decision record."
    )
)

Training and evaluation data

Dataset: s3nh/alpaca-dolly-instruction-only-polish

Each row has been converted into conversation using this function:

system_message = """Jesteś przyjaznym chatbotem"""

def create_conversation(sample) -> dict:
    strip_characters = "\"'"
    return {
        "messages": [
            {"role": "system", "content": system_message},
            {"role": "user",
             "content": f"{sample['instruction'].strip(strip_characters)} "
                        f"{sample['input'].strip(strip_characters)}"},
            {"role": "assistant",
             "content": f"{sample['output'].strip(strip_characters)}"}
        ]
    }

Train/test split: 90%/10%

Training procedure

GPU: 2x RTX 4060Ti 16GB Training time: ~1 hour

Using accelerate + deepspeed with config:

compute_environment: LOCAL_MACHINE
debug: false
deepspeed_config:
  gradient_accumulation_steps: 2
  zero3_init_flag: false
  zero_stage: 1
distributed_type: DEEPSPEED
downcast_bf16: 'no'
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 2
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

Training hyperparameters

Lora config:

peft_config = LoraConfig(
    lora_alpha=128,
    lora_dropout=0.05,
    r=256,
    bias="none",
    target_modules="all-linear",
    task_type="CAUSAL_LM"
)

Training arguments:

args = TrainingArguments(
    output_dir="Qra-1b-dolly-instruction-0.1",
    num_train_epochs=3,
    per_device_train_batch_size=3,
    gradient_accumulation_steps=2,
    gradient_checkpointing=True,
    optim="adamw_torch_fused",
    logging_steps=10,
    save_strategy="epoch",
    learning_rate=2e-4,
    bf16=True,
    tf32=True,
    max_grad_norm=0.3,
    warmup_ratio=0.03,
    lr_scheduler_type="constant",
    push_to_hub=False,
    report_to=["tensorboard"],
)

Framework versions

• PEFT 0.10.0
• Transformers 4.39.2
• Pytorch 2.2.2+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2