Ramikan-BR/tinyllama-coder-py-4bit-v10

Uploaded model

• Developed by: Ramikan-BR
• Model type: [text-generation/Python Coder]
• Language(s) (NLP): [en]
• License: apache-2.0
• Finetuned from model : unsloth/tinyllama-chat-bnb-4bit

Model Description

Training Data

datasets: Ramikan-BR/data-oss_instruct-decontaminated_python.jsonl

Training Procedure

The model was refined using Unsloath. The dataset ise-uiuc/Magicoder-OSS-Instruct-75K was adjusted, leaving only data on python and divided into 10 parts, each refinement occurred for 2 epochs, using adafactor optimizer or adamw_8bit (adafactor seems to deliver less loss).

Model Sources [optional]

base_model: unsloth/tinyllama-chat-bnb-4bit

model: Ramikan-BR/tinyllama-coder-py-4bit-v10 gguf_f16: tinyllama-coder-py-4bit-v10-unsloth.F16.gguf gguf_Q4_K_M: tinyllama-coder-py-4bit-v10-unsloth.Q4_K_M.gguf gguf_Q8_0: tinyllama-coder-py-4bit-v10-unsloth.Q8_0.gguf

Training Hyperparameters

Notebook Unsloath that I used for AI refinement: TinyLlama

%%capture
# Installs Unsloth, Xformers (Flash Attention) and all other packages!
!pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
!pip install --no-deps xformers trl peft accelerate bitsandbytes # xformers "xformers<0.0.26"

import os
from google.colab import drive
drive.mount('/content/drive')

from unsloth import FastLanguageModel
import torch
max_seq_length = 4096 # Choose any! We auto support RoPE Scaling internally!
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.

# 4bit pre quantized models we support for 4x faster downloading + no OOMs.
fourbit_models = [
    "unsloth/mistral-7b-bnb-4bit",
    "unsloth/mistral-7b-instruct-v0.2-bnb-4bit",
    "unsloth/llama-2-7b-bnb-4bit",
    "unsloth/llama-2-13b-bnb-4bit",
    "unsloth/codellama-34b-bnb-4bit",
    "unsloth/tinyllama-bnb-4bit",
    "unsloth/gemma-7b-bnb-4bit", # New Google 6 trillion tokens model 2.5x faster!
    "unsloth/gemma-2b-bnb-4bit",
] # More models at https://huggingface.co./unsloth

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Ramikan-BR/tinyllama-coder-py-4bit_LORA-v9", # "unsloth/tinyllama" for 16bit loading
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
    # token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
)

model = FastLanguageModel.get_peft_model(
    model,
    r = 256, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 512,
    lora_dropout = 0, # Currently only supports dropout = 0
    bias = "none",    # Currently only supports bias = "none"
    use_gradient_checkpointing = True, # @@@ IF YOU GET OUT OF MEMORY - set to True @@@
    random_state = 3407,
    use_rslora = False,  # We support rank stabilized LoRA
    loftq_config = None, # And LoftQ
)

alpaca_prompt = """Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Input:
{}

### Output:
{}"""

EOS_TOKEN = tokenizer.eos_token
def formatting_prompts_func(examples):
    inputs       = examples["problem"]
    outputs      = examples["solution"]
    texts = []
    for input, output in zip(inputs, outputs):
        # Must add EOS_TOKEN, otherwise your generation will go on forever!
        text = alpaca_prompt.format(input, output) + EOS_TOKEN
        texts.append(text)
    return { "text" : texts}
pass

from datasets import load_dataset
dataset = load_dataset('json', data_files='/content/drive/MyDrive/data-oss_instruct-py-10.jsonl', split='train')
dataset = dataset.map(formatting_prompts_func, batched=True)

from trl import SFTTrainer
from transformers import TrainingArguments
from unsloth import is_bfloat16_supported
from transformers.utils import logging
logging.set_verbosity_info()

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    dataset_text_field = "text",
    max_seq_length = max_seq_length,
    dataset_num_proc = 2,
    packing = True, # Packs short sequences together to save time!
    args = TrainingArguments(
        per_device_train_batch_size = 2,
        gradient_accumulation_steps = 256,
        warmup_ratio = 0.1,
        num_train_epochs = 2,
        learning_rate = 2e-4,
        fp16 = not torch.cuda.is_bf16_supported(),
        bf16 = torch.cuda.is_bf16_supported(),
        logging_steps = 1,
        optim = "adafactor", # adamw_torch ou adamw_torch_fused +10% velocidade ou adafactor ou adamw_8bit
        weight_decay = 0.1,
        lr_scheduler_type = "linear",
        seed = 3407,
        output_dir = "outputs",
    ),
)

trainer_stats = trainer.train()

model.save_pretrained("lora_model") # Local saving
tokenizer.save_pretrained("lora_model")
model.push_to_hub("Ramikan-BR/tinyllama-coder-py-4bit_LORA-v10", token = "hf_...") # Online saving
tokenizer.push_to_hub("Ramikan-BR/tinyllama-coder-py-4bit_LORA-v10", token = "hf_...") # Online saving

# Merge to 16bit
model.save_pretrained_merged("model", tokenizer, save_method = "merged_16bit",)
model.push_to_hub_merged("Ramikan-BR/tinyllama-coder-py-4bit-v10", tokenizer, save_method = "merged_16bit", token = "hf_...")

# Merge to 4bit
if False: model.save_pretrained_merged("model", tokenizer, save_method = "merged_4bit",)
if False: model.push_to_hub_merged("Ramikan-BR/tinyllama-coder-py-4bit-v10", tokenizer, save_method = "merged_4bit", token = "hf_...")

# Just LoRA adapters
if False: model.save_pretrained_merged("model", tokenizer, save_method = "lora",)
if False: model.push_to_hub_merged("Ramikan-BR/tinyllama-coder-py-4bit-v10", tokenizer, save_method = "lora", token = "hf_...")

# Save to 8bit Q8_0
model.save_pretrained_gguf("model", tokenizer,)
model.push_to_hub_gguf("Ramikan-BR/tinyllama-coder-py-4bit-v10", tokenizer, token = "hf_...")

# Save to 16bit GGUF
model.save_pretrained_gguf("model", tokenizer, quantization_method = "f16")
model.push_to_hub_gguf("Ramikan-BR/tinyllama-coder-py-4bit-v10", tokenizer, quantization_method = "f16", token = "hf_...")

# Save to q4_k_m GGUF
model.save_pretrained_gguf("model", tokenizer, quantization_method = "q4_k_m")
model.push_to_hub_gguf("Ramikan-BR/tinyllama-coder-py-4bit-v10", tokenizer, quantization_method = "q4_k_m", token = "hf_...")

Loss for 5 epochs in the last training session of the last part of the dataset:
==((====))==  Unsloth - 2x faster free finetuning | Num GPUs = 1
   \\   /|    Num examples = 407 | Num Epochs = 5
O^O/ \_/ \    Batch size per device = 2 | Gradient Accumulation steps = 256
\        /    Total batch size = 512 | Total steps = 5
 "-____-"     Number of trainable parameters = 201,850,880
 [5/5 29:36, Epoch 3/5]
Step	Training Loss
1	0.568000
2	0.145300
3	0.506100
4	0.331900
5	0.276100

Quick test 1 after training the last part of the dataset:

# alpaca_prompt = Copied from above
FastLanguageModel.for_inference(model) # Enable native 2x faster inference
inputs = tokenizer(
[
    alpaca_prompt.format(
        "Continue the fibonnaci sequence.", # instruction
        "1, 1, 2, 3, 5, 8", # input
        "", # output - leave this blank for generation!
    )
], return_tensors = "pt").to("cuda")

AI Response: ['<s> Below is an instruction that describes a task. Write a response that appropriately completes the request.\n### Input:\nContinue the fibonnaci sequence.\n\n### Output:\n1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 420, 787, 1444, 2881, 4765, 8640']

Quick test 2 after training the last part of the dataset:

# alpaca_prompt = Copied from above
FastLanguageModel.for_inference(model) # Enable native 2x faster inference
inputs = tokenizer(
[
    alpaca_prompt.format(
        "Continue the fibonnaci sequence.", # instruction
        "1, 1, 2, 3, 5, 8", # input
        "", # output - leave this blank for generation!
    )
], return_tensors = "pt").to("cuda")

from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 128)

AI Response: <s> Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Input:
Continue the fibonnaci sequence.

### Output:
1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 420, 787, 1444, 2881, 4765, 8640, 17281, 31362, 65325, 128672, 251345, 410000, 720000, 1280000,

Quick test 3 after training the last part of the dataset:

if False:
    from unsloth import FastLanguageModel
    model, tokenizer = FastLanguageModel.from_pretrained(
        model_name = "lora_model", # YOUR MODEL YOU USED FOR TRAINING
        max_seq_length = max_seq_length,
        dtype = dtype,
        load_in_4bit = load_in_4bit,
    )
    FastLanguageModel.for_inference(model) # Enable native 2x faster inference

# alpaca_prompt = You MUST copy from above!

inputs = tokenizer(
[
    alpaca_prompt.format(
        "What is a famous tall tower in Paris?", # instruction
        "", # input
        "", # output - leave this blank for generation!
    )
], return_tensors = "pt").to("cuda")

from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 64)

AI Response: <s> Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Input:
What is a famous tall tower in Paris?

### Output:
The famous tall tower in Paris is the Eiffel Tower. It is a 300-meter-tall steel tower located in the heart of Paris, France. The tower was built in 18892 and is a popular tourist attraction. It is also a symbol of the city

outputs = model.generate(**inputs, max_new_tokens = 64, use_cache = True)
tokenizer.batch_decode(outputs)

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