[!NOTE]
The ST format is converted from PTH, designed for use with RWKV webgpu inference engines such as Ai00 and web-rwkv.
1️⃣ What are G0 / G1 / G1a2 / G1b?
The labels G0a / G1a / G1a2, etc., indicate different versions of the training data.
Data quality ranking: G1b > G1a3 > G1a2 > G1a > G1 > G0a2 > G0.
The RWKV7-G1a model is an advanced version of RWKV7-G1, further trained on 1T of high-quality inference and instruction data.
Future updates will continue to add more high-quality data to form the G1b dataset, along with the release of the RWKV7-G1b model series.
2️⃣ How to choose the best model?
Check the date in the model name.
For models with the same parameter size, newer models are better!
For example, a 2.9B model released on 251005 will definitely outperform a version released on 250606.
3️⃣ Prompt Format
The RWKV7-G series supports both no-think-mode and think-mode.
no-think-mode Chat prompt
[!TIP]
better replace all \n\n in USER_PROMPT to \n as i am using \n\n as "chat round separator" in pretrain data)
