Model Card

Summary

This model was trained using H2O LLM Studio.

Base model: h2oai/h2o-danube3-500m-chat
Fine-tuning dataset: zakariarada/oasst

Training

To train the model using your custom dataset, you can follow the steps below. This example demonstrates how to fine-tune the h2oai/h2o-danube3-500m-chat model using the Hugging Face transformers library.

Code Example

import pandas as pd
from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    TrainingArguments,
    Trainer
)
from datasets import Dataset

# Load Dataset
data_path = "train_full.pq"
df = pd.read_parquet(data_path)

# Prepare Dataset for Training
dataset = Dataset.from_pandas(df)

def preprocess_function(examples):
    # Combine 'instruction' and 'parent_id' as input prompt
    instruction = examples["instruction"]
    parent_id = examples["parent_id"]
    input_prompt = f"{parent_id}: {instruction}" if parent_id else instruction
    return {
        "input_text": input_prompt,
        "target_text": examples["output"]
    }

# Preprocess Dataset
dataset = dataset.map(preprocess_function, remove_columns=["id", "parent_id", "instruction", "output"])

# Load Tokenizer and Model
model_name = "h2oai/h2o-danube3-500m-chat"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Tokenize Data
def tokenize_function(examples):
    return tokenizer(
        examples["input_text"],
        padding="max_length",
        truncation=True,
        max_length=512
    )

tokenized_dataset = dataset.map(tokenize_function, batched=True)

# Training Arguments
training_args = TrainingArguments(
    output_dir="./output/TCLM-beta/",  # Directory to save model checkpoints
    num_train_epochs=3,  # Increase epochs for better fine-tuning results
    per_device_train_batch_size=4,  # Adjust based on GPU memory, increase if possible
    gradient_accumulation_steps=4,  # Accumulate gradients to simulate a larger batch size
    evaluation_strategy="steps",  # Evaluate more frequently for detailed tracking
    eval_steps=500,  # Evaluate every 500 steps to track progress without over-evaluating
    save_strategy="steps",  # Save checkpoints during training
    save_steps=500,  # Save model every 500 steps
    save_total_limit=2,  # Limit to the two best models to save disk space
    learning_rate=5e-5,  # Lower learning rate for fine-tuning
    weight_decay=0.01,  # Slight weight decay to prevent overfitting
    lr_scheduler_type="cosine",  # Cosine schedule for smoother learning rate decay
    warmup_ratio=0.06,  # Warmup to stabilize initial training
    logging_dir="./logs",  # Directory to save training logs
    logging_steps=50,  # Log progress every 50 steps for better monitoring
    fp16=True,  # Enable mixed precision for faster training with less memory
    load_best_model_at_end=True,  # Load the best model at the end based on evaluation metric
    metric_for_best_model="eval_loss",  # Use evaluation loss to determine the best model
    greater_is_better=False,  # Lower loss is better
)


# Trainer Setup
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset,
    tokenizer=tokenizer,
)

# Train Model
trainer.train()

Usage

To use the model with the transformers library on a machine with GPUs, first make sure you have the transformers library installed.

pip install transformers==4.45.0

Also make sure you are providing your huggingface token to the pipeline if the model is lying in a private repo.

Either leave token=True in the pipeline and login to hugginface_hub by running

import huggingface_hub
huggingface_hub.login(<ACCESS_TOKEN>)

Or directly pass your to token in the pipeline

from transformers import pipeline

generate_text = pipeline(
    model="zakariarada/TCLM-beta",
    torch_dtype="auto",
    trust_remote_code=True,
    device_map={"": "cuda:0"},
    token=True,
)

# generate configuration can be modified to your needs
# generate_text.model.generation_config.min_new_tokens = 2
# generate_text.model.generation_config.max_new_tokens = 256
# generate_text.model.generation_config.do_sample = False
# generate_text.model.generation_config.num_beams = 1
# generate_text.model.generation_config.temperature = float(0.0)
# generate_text.model.generation_config.repetition_penalty = float(1.0)

messages = [
    {"role": "user", "content": "Hi, how are you?"},
    {"role": "assistant", "content": "I'm doing great, how about you?"},
    {"role": "user", "content": "Why is drinking water so healthy?"},
]

res = generate_text(
    messages,
    renormalize_logits=True
)
print(res[0]["generated_text"][-1]['content'])

You can print a sample prompt after applying chat template to see how it is feed to the tokenizer:

print(generate_text.tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
))

You may also construct the pipeline from the loaded model and tokenizer yourself and consider the preprocessing steps:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "zakariarada/TCLM-beta"  # either local folder or Hugging Face model name
# Important: The prompt needs to be in the same format the model was trained with.
# You can find an example prompt in the experiment logs.
messages = [
    {"role": "user", "content": "Hi, how are you?"},
    {"role": "assistant", "content": "I'm doing great, how about you?"},
    {"role": "user", "content": "Why is drinking water so healthy?"},
]

tokenizer = AutoTokenizer.from_pretrained(
    model_name,
    trust_remote_code=True,
)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map={"": "cuda:0"},
    trust_remote_code=True,
)
model.cuda().eval()

# generate configuration can be modified to your needs
# model.generation_config.min_new_tokens = 2
# model.generation_config.max_new_tokens = 256
# model.generation_config.do_sample = False
# model.generation_config.num_beams = 1
# model.generation_config.temperature = float(0.0)
# model.generation_config.repetition_penalty = float(1.0)

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
).to("cuda")

tokens = model.generate(
    input_ids=inputs["input_ids"],
    attention_mask=inputs["attention_mask"],
    renormalize_logits=True
)[0]

tokens = tokens[inputs["input_ids"].shape[1]:]
answer = tokenizer.decode(tokens, skip_special_tokens=True)
print(answer)

Quantization and sharding

You can load the models using quantization by specifying load_in_8bit=True or load_in_4bit=True. Also, sharding on multiple GPUs is possible by setting device_map=auto.

Model Architecture

LlamaForCausalLM(
  (model): LlamaModel(
    (embed_tokens): Embedding(32000, 1536, padding_idx=0)
    (layers): ModuleList(
      (0-15): 16 x LlamaDecoderLayer(
        (self_attn): LlamaSdpaAttention(
          (q_proj): Linear(in_features=1536, out_features=1536, bias=False)
          (k_proj): Linear(in_features=1536, out_features=768, bias=False)
          (v_proj): Linear(in_features=1536, out_features=768, bias=False)
          (o_proj): Linear(in_features=1536, out_features=1536, bias=False)
          (rotary_emb): LlamaRotaryEmbedding()
        )
        (mlp): LlamaMLP(
          (gate_proj): Linear(in_features=1536, out_features=4096, bias=False)
          (up_proj): Linear(in_features=1536, out_features=4096, bias=False)
          (down_proj): Linear(in_features=4096, out_features=1536, bias=False)
          (act_fn): SiLU()
        )
        (input_layernorm): LlamaRMSNorm((1536,), eps=1e-05)
        (post_attention_layernorm): LlamaRMSNorm((1536,), eps=1e-05)
      )
    )
    (norm): LlamaRMSNorm((1536,), eps=1e-05)
    (rotary_emb): LlamaRotaryEmbedding()
  )
  (lm_head): Linear(in_features=1536, out_features=32000, bias=False)
)

Model Configuration

This model was trained using H2O LLM Studio and with the configuration in cfg.yaml. Visit H2O LLM Studio to learn how to train your own large language models.

Disclaimer

Please read this disclaimer carefully before using the large language model provided in this repository. Your use of the model signifies your agreement to the following terms and conditions.

Biases and Offensiveness: The large language model is trained on a diverse range of internet text data, which may contain biased, racist, offensive, or otherwise inappropriate content. By using this model, you acknowledge and accept that the generated content may sometimes exhibit biases or produce content that is offensive or inappropriate. The developers of this repository do not endorse, support, or promote any such content or viewpoints.
Limitations: The large language model is an AI-based tool and not a human. It may produce incorrect, nonsensical, or irrelevant responses. It is the user's responsibility to critically evaluate the generated content and use it at their discretion.
Use at Your Own Risk: Users of this large language model must assume full responsibility for any consequences that may arise from their use of the tool. The developers and contributors of this repository shall not be held liable for any damages, losses, or harm resulting from the use or misuse of the provided model.
Ethical Considerations: Users are encouraged to use the large language model responsibly and ethically. By using this model, you agree not to use it for purposes that promote hate speech, discrimination, harassment, or any form of illegal or harmful activities.
Reporting Issues: If you encounter any biased, offensive, or otherwise inappropriate content generated by the large language model, please report it to the repository maintainers through the provided channels. Your feedback will help improve the model and mitigate potential issues.
Changes to this Disclaimer: The developers of this repository reserve the right to modify or update this disclaimer at any time without prior notice. It is the user's responsibility to periodically review the disclaimer to stay informed about any changes.

By using the large language model provided in this repository, you agree to accept and comply with the terms and conditions outlined in this disclaimer. If you do not agree with any part of this disclaimer, you should refrain from using the model and any content generated by it.