Đà mã 2 (Llama2 architecture)

Dama2 is an autoregressive Large Language Model (LLM), based on Llama2's model architecture. Dama2 was trained on part of the Common Crawl dataset in Vietnamese and English.

Details will be available soon.

To contact us, mail to: leanhcuong@gmail.com (Lê Anh Cường) | hieunguyen1053@outlook.com (Hiếu)

How to use

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("vietgpt/dama-2-7b")
model = AutoModelForCausalLM.from_pretrained("vietgpt/dama-2-7b", low_cpu_mem_usage=True)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 
model.to(device)

prompt = "Địa chỉ trường Đại học Tôn Đức Thắng nằm ở số"
input_ids = tokenizer(prompt, return_tensors="pt")['input_ids'].to(device)

gen_tokens = model.generate(input_ids, max_length=max_length, repetition_penalty=1.1)

print(tokenizer.batch_decode(gen_tokens)[0])

{
  "results": {
    "lambada_vi": {
      "ppl": 17.662483545322115,
      "ppl_stderr": 0.46441057543941494,
      "acc": 0.34159672067148156,
      "acc_stderr": 0.004685401990271572
    }
  },
  "versions": {
    "lambada_vi": null
  },
  "config": {
    "model": "hf-causal",
    "model_args": "pretrained=vietgpt/dama-2-7b",
    "num_fewshot": 0,
    "batch_size": null,
    "batch_sizes": [],
    "device": "cuda:1",
    "no_cache": false,
    "limit": null,
    "bootstrap_iters": 100000,
    "description_dict": {}
  }
}
hf-causal (pretrained=vietgpt/dama-2-7b), limit: None, provide_description: False, num_fewshot: 0, batch_size: None
|   Task   |Version|Metric| Value |   |Stderr|
|----------|-------|------|------:|---|-----:|
|lambada_vi|       |ppl   |17.6625|±  |0.4644|
|          |       |acc   | 0.3416|±  |0.0047|