fla-hub / rwkv7-168M-pile

rwkv7-168M-pile

This is RWKV-7 model under flash-linear attention format.

Warning! Currently training is not tested. Only inference is supported.

Model Details

Model Description

• Developed by: Bo Peng, Yu Zhang, Songlin Yang, Ruochong Zhang
• Funded by: Shenzhen Yuanshi Intelligent Co. Ltd.
• Model type: RWKV7
• Language(s) (NLP): English
• License: Apache-2.0
• Parameter count: 165M
• Tokenizer: GPT-NeoX 20B tokenizer

Model Sources

• Repository: https://github.com/fla-org/flash-linear-attention ; https://github.com/BlinkDL/RWKV-LM
• Paper: With in Progress
• Weights: Converted from https://modelscope.cn/models/RWKV/rwkv-7-pile/file/view/master?fileName=RWKV-x070-Pile-168M-20241120-ctx4096.pth

Uses

Install flash-linear-attention and the latest version of transformers before using this model:

pip install git+https://github.com/fla-org/flash-linear-attention
pip install 'transformers>=4.48.0'

Direct Use

You can use this model just as any other HuggingFace models:

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained('fla-hub/rwkv7-168M-pile', trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained('fla-hub/rwkv7-168M-pile', trust_remote_code=True)

Training Details

Training Data

This model is trained on the Pile with a total of 332 billion tokens.

Training Hyperparameters

• Training regime: bfloat16, lr 8e-4 to 3e-5 cosine decay, wd 0.1, bsz 8x30x4096

Evaluation

Metrics

lambada_openai : ppl 14.2 acc 45.6%

piqa : acc 65.5%