The fields like G0a / G1a / G1a2 in RWKV model names indicate versions of the training data. In terms of data quality, the ranking is: G1b > G1a3 > G1a2 > G1a > G1 > G0a2 > G0.
The RWKV7-G1a model is an advanced version of RWKV7-G1 that was further trained with 1T (1 trillion tokens) of high-quality inference and instruction data. RWKV7-G1a2 was produced by continuing to add more data and training on top of RWKV7-G1a. And so on.
[!TIP]
More high-quality data will be added later to form the G1b dataset, and RWKV7-G1b series models will also be trained and open-sourced.
What is the difference between the RWKV7-G series and the World series?
The RWKV7-G series supports an inference mode, which can be activated using the following format:
User: USER_PROMPT
Assistant: <think
How to choose the best model?
Look at the date in the model name — for the same parameter size, a newer model is better!
For example, for the same 1.5B model, a G1a2 version released on 251005 will definitely be superior to a G1 version released on 250429.
[!WARNING]
For the 0.1B and 0.4B models, we recommend using FP16/Q8_0 quantization. Otherwise, the models may fail to complete tasks due to precision loss caused by quantization.
