h2o-danube3

alt="Slightly cropped image" style="width: 102%; height: 102%; object-fit: cover; object-position: center; margin: -5% -5% -5% -5%;">

Summary

h2o-danube3-500m-chat is a chat fine-tuned model by H2O.ai with 500 million parameters. We release two versions of this model:

This model was trained using H2O LLM Studio.

Can be run natively and fully offline on phones - try it yourself with H2O AI Personal GPT.

Model Architecture

We adjust the Llama 2 architecture for a total of around 500m parameters. For details, please refer to our Technical Report. We use the Mistral tokenizer with a vocabulary size of 32,000 and train our model up to a context length of 8,192.

The details of the model architecture are:

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>=4.42.3

import torch
from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="h2oai/h2o-danube3-500m-chat",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

# We use the HF Tokenizer chat template to format each message
# https://huggingface.co/docs/transformers/main/en/chat_templating
messages = [
    {"role": "user", "content": "Why is drinking water so healthy?"},
]
prompt = pipe.tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
res = pipe(
    prompt,
    return_full_text=False,
    max_new_tokens=256,
)
print(res[0]["generated_text"])

This will apply and run the correct prompt format out of the box:

Alternatively, one can also run it via:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "h2oai/h2o-danube3-500m-chat"

tokenizer = AutoTokenizer.from_pretrained(
    model_name,
)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {"role": "user", "content": "Why is drinking water so healthy?"},
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(
    prompt, return_tensors="pt", add_special_tokens=False
).to("cuda")

# generate configuration can be modified to your needs
tokens = model.generate(
    input_ids=inputs["input_ids"],
    attention_mask=inputs["attention_mask"],
    min_new_tokens=2,
    max_new_tokens=256,
)[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()
        (post_attention_layernorm): LlamaRMSNorm()
      )
    )
    (norm): LlamaRMSNorm()
  )
  (lm_head): Linear(in_features=1536, out_features=32000, bias=False)
)

Benchmarks

🤗 Open LLM Leaderboard v1

MT-Bench

First Turn: 4.16
Second Turn: 2.40
Average: 3.28

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