llama3-8B-slerp-biomed-chat-chinese
llama3-8B-slerp-biomed-chat-chinese is a merge of the following models using LazyMergekit:
𧩠Configuration
slices:
- sources:
- model: shanchen/llama3-8B-slerp-med-chinese
layer_range: [0,32]
- model: shenzhi-wang/Llama3-8B-Chinese-Chat
layer_range: [0,32]
merge_method: slerp
base_model: shenzhi-wang/Llama3-8B-Chinese-Chat
parameters:
t:
- filter: self_attn
value: [0.3, 0.5, 0.5, 0.7, 1]
- filter: mlp
value: [1, 0.7, 0.5, 0.5, 0.3]
- value: 0.5
dtype: bfloat16
π» Usage
!pip install -qU transformers accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM
model_id = "shanchen/llama3-8B-slerp-biomed-chat-chinese"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id, torch_dtype="auto", device_map="auto"
)
messages = [
{"role": "user", "content": "Can you speak Japanese?"},
]
input_ids = tokenizer.apply_chat_template(
messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
outputs = model.generate(
input_ids,
max_new_tokens=192 max#8192,
do_sample=True,
temperature=0.6,
top_p=0.9,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))
- Downloads last month
- 2,644
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 shanchen/llama3-8B-slerp-biomed-chat-chinese
Merge model
this model