Edit model card

Релиз вихря 0.5*

Долили сильно больше данных в sft, теперь стабильнее работает json и multiturn, слегка подточили параметры претрена модели, добавили RoPE на 32к контекста

Added a lot more data to sft, now json and multiturn work more stable on long context and hard prompts



from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model = AutoModelForCausalLM.from_pretrained("Vikhrmodels/it-5.3-fp16-32k",
                                             device_map="auto",
                                             attn_implementation="sdpa",
                                             torch_dtype=torch.bfloat16)

tokenizer = AutoTokenizer.from_pretrained("Vikhrmodels/it-5.3-fp16-32k")
from transformers import  AutoTokenizer, pipeline
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
prompts = [
    "В чем разница между фруктом и овощем?",
    "Годы жизни колмагорова?"]

def test_inference(prompt):
    prompt = pipe.tokenizer.apply_chat_template([{"role": "user", "content": prompt}], tokenize=False, add_generation_prompt=True)
    print(prompt)
    outputs = pipe(prompt, max_new_tokens=512, do_sample=True, num_beams=1, temperature=0.25, top_k=50, top_p=0.98, eos_token_id=79097)
    return outputs[0]['generated_text'][len(prompt):].strip()


for prompt in prompts:
    print(f"    prompt:\n{prompt}")
    print(f"    response:\n{test_inference(prompt)}")
    print("-"*50)

@article{nikolich2024vikhr,
  title={Vikhr: The Family of Open-Source Instruction-Tuned Large Language Models for Russian},
  author={Aleksandr Nikolich and Konstantin Korolev and Artem Shelmanov},
  journal={arXiv preprint arXiv:2405.13929},
  year={2024},
  url={https://arxiv.org/pdf/2405.13929}
}
Downloads last month
28
Safetensors
Model size
7.63B params
Tensor type
FP16
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for Vikhrmodels/it-5.3-fp16-32k

Finetunes
2 models
Quantizations
2 models