L-MChat-7b

L-MChat-7b is a merge of the following models:

• Nexusflow/Starling-LM-7B-beta
• FuseAI/FuseChat-7B-VaRM

Configuration

slices:
  - sources:
      - model: Nexusflow/Starling-LM-7B-beta
        layer_range: [0, 32]
      - model: FuseAI/FuseChat-7B-VaRM
        layer_range: [0, 32]
merge_method: slerp
base_model: FuseAI/FuseChat-7B-VaRM
parameters:
  t:
    - filter: self_attn
      value: [0, 0.5, 0.3, 0.7, 1]
    - filter: mlp
      value: [1, 0.5, 0.7, 0.3, 0]
    - value: 0.5
dtype: bfloat16

Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "Artples/M-LChat-7b"
messages = [{"role": "user", "content": "What is a large language model?"}]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])

License

Apache 2.0 but you cannot use this model to directly compete with OpenAI.

How?

Usage of LazyMergekit.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	69.57
AI2 Reasoning Challenge (25-Shot)	65.61
HellaSwag (10-Shot)	84.59
MMLU (5-Shot)	65.44
TruthfulQA (0-shot)	50.94
Winogrande (5-shot)	81.37
GSM8k (5-shot)	69.45