TimeMobius's picture
Update README.md
79168c1
metadata
license: apache-2.0
language:
  - fr
  - it
  - de
  - es
  - en
  - zh
inference: false

Model Card for Mobius-12B-base-m1

The Mobius-12B-base-m1 Large Language Model (LLM) is a pretrained model based on RWKV v5 arch. We utilized 0.01 billion tokens to conduct post-training on this model for alignment benchmarks, excluding the utilization of DPO and SFT. The process took approximately 10 hours, employing 4 * a800.

Warning

This repo contains weights that are not compatible with Hugging Face transformers library yet. But you can try thisPR as well. RWKV runner or AI00 server also work.

Instruction|Chat format

This format must be strictly respected, otherwise the model will generate sub-optimal outputs.

The template used to build a prompt for the Instruct model is defined as follows:

User: {Instruction|prompt}\n\nAssistant:

Run the model

need to convert checkpoint to HF format

Need to install this PR pip install -e git://github.com/BBuf/transformers.git

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("TimeMobius/Mobius-12B-base-m1", torch_dtype=torch.float16).to(0)
tokenizer = AutoTokenizer.from_pretrained("TimeMobius/Mobius-12B-base-m1", trust_remote_code=True)

text = "x"
prompt = f'Question: {text.strip()}\n\nAnswer:'

inputs = tokenizer(prompt, return_tensors="pt").to(0)
output = model.generate(inputs["input_ids"], max_new_tokens=40)
print(tokenizer.decode(output[0].tolist(), skip_special_tokens=True))

Limitations

The Mobius base m1 is the base model can be easily fine-tuned to achieve compelling performance. if you wanna better benchmark results use DPO and SFT ,details in readme

Benchmark

Mobius-12B-base-m1
lambda ppl 3.41
lambda 0.72
piqa 0.78
hellaswag 10 shots 0.72
winogrande 0.68
arc_challenge 25shots 0.47
arc_easy 0.73
openbookqa 0.40
sciq 0.93

@TimeMobius