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.
2️⃣ How to Choose the Best Model?
Look at the Date in the Model Name—with the same parameters, the newer the model, the better!
For instance, for the 1.5B models, the G1a2 version released on 251005 is definitely superior to the G1 version released on 250429.
3️⃣ 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.
