README.md

metadata

license: apache-2.0
datasets:
  - IlyaGusev/gazeta
  - IlyaGusev/ru_turbo_alpaca_evol_instruct
  - IlyaGusev/ru_turbo_alpaca
  - IlyaGusev/ru_turbo_saiga
  - RussianNLP/russian_super_glue
language:
  - ru
pipeline_tag: question-answering

The model was trained on part of the datasets

IlyaGusev/gazeta ,

IlyaGusev/ru_turbo_alpaca_evol_instruct,

IlyaGusev/ru_turbo_alpaca,

IlyaGusev/ru_turbo_saiga ,

RussianNLP/russian_super_glue (muserc)

using LoRA

Base_model NousResearch/Yarn-Llama-2-7b-64k

Need cuda > 11.4

GPU A100

  
!pip install peft
!pip install flash-attn --no-build-isolation
!pip install git+https://github.com/HazyResearch/flash-attention.git#subdirectory=csrc/rotary

model = AutoModelForCausalLM.from_pretrained(
            'geldarr/saiga-Yarn-Llama-2-7b-64k',
            trust_remote_code=True,
            torch_dtype=torch.float16,
            device_map={'':0}
        )
tokenizer = AutoTokenizer.from_pretrained('geldarr/saiga-Yarn-Llama-2-7b-64k', use_fast=False)

big_prompts = '''<s>system\nТы — Сайга, русскоязычный автоматический ассистент. Ты разговариваешь с людьми и помогаешь им.</s>\n
<s>user
Дай ответы на вопрос основываясь только на тексте ниже:\n
вопрос?

Текст <65536 tokens

</s>
<s>bot
'''
```python
gen_config = {
        "pad_token_id": 0,
        "bos_token_id": 1,
        "eos_token_id": 2,
        "temperature": 0.4,
        "top_p": 0.9,
        "top_k": 50,
        "do_sample": True,
        "max_new_tokens": 15360,
        "repetition_penalty": 1.1,
        "no_repeat_ngram_size": 15,
    }
generation_config = GenerationConfig.from_dict(gen_config)

def generate(model, tokenizer, prompt, generation_config):
    data = tokenizer(prompt, return_tensors="pt")

    data = {k: v.to(model.device) for k, v in data.items()}
    output_ids = model.generate(
        **data,
        generation_config=generation_config
    )[0]
    output_ids = output_ids[len(data["input_ids"][0]):]
    output = tokenizer.decode(output_ids)
    return output.strip()

output = generate(model, tokenizer, big_prompts, generation_config)

print(output)