shoumenchougou/RWKV-ST-model

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

better replace all \n\n in USER_PROMPT to \n as i am using \n\n as "chat round separator" in pretrain data)

System: YOU_CAN_USE_SYSTEM_IF_NEEDED

User: PREVIOUS_STUFF

Assistant: PREVIOUS_STUFF

User: USER_PROMPT

Assistant:

think-mode prompt:

User: USER_PROMPT

Assistant: <think

---

Think prompt, alternative style output, valid for 20250922 and newer models. Note there is a space before the "think" after USER_PROMPT:

User: USER_PROMPT think

Assistant: <think

Shorter think (think a bit), same style:

User: USER_PROMPT think a bit

Assistant: <think

Longer think (think a lot), same style:

User: USER_PROMPT think a lot

Assistant: <think

---

Fake think prompt:

User: USER_PROMPT

Assistant: <think>
</think