Seeking Improvements and Configuration Advice for Longer Responses and Larger Tokens
I have experimented with the model up until version 7. However, I have been unable to obtain lengthy responses despite various adjustments. Are there any plans to enhance this aspect of the model, or could you suggest specific settings that might help in achieving longer responses?
Additionally, could you provide some insights into the model's capability to handle larger tokens (exceeding 2048)?
You can simply talk to Alice in ChatRWKV v2 (Raven v8) and get very long responses.
The 7B and 14B Raven models can handle ctxlen 4096 and beyond.
I've verified that the model generates longer content in versions 8 and above, which is great to see. However, one major weakness of these models is their inability to consistently follow clear instructions. For instance, when asked to 'List 20 breeds of dogs', the model often generates a list with more than 20 dogs and with several duplicates, regardless of the settings used. While I haven't tested the newer v9 model yet, I suspect it may face similar limitations. Could this be a limitation with 100% RNN models?
Reduce top_p to 0 or 0.1 or 0.2 or 0.5