rixty/SillyTavern
1
0
Fork 0
mirror of https://github.com/SillyTavern/SillyTavern.git synced 2025-06-05 21:59:27 +02:00
Code Issues Packages Projects Releases Wiki Activity

Add extra padding for 0301 turbo model.

Browse Source
This commit is contained in:
Cohee
2023-10-16 01:21:14 +03:00
parent 5b954a3584
commit 02dca52948
1 changed files with 8 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/tokenizers.js
View File

@ -288,7 +288,8 @@ function registerEndpoints(app, jsonParser) {
if (!req.body) return res.sendStatus(400); if (!req.body) return res.sendStatus(400);
let num_tokens = 0; let num_tokens = 0;
const model = getTokenizerModel(String(req.query.model || '')); const queryModel = String(req.query.model || '');
const model = getTokenizerModel(queryModel);
if (model == 'claude') { if (model == 'claude') {
num_tokens = countClaudeTokens(claude_tokenizer, req.body); num_tokens = countClaudeTokens(claude_tokenizer, req.body);
@ -316,6 +317,12 @@ function registerEndpoints(app, jsonParser) {
} }
num_tokens += tokensPadding; num_tokens += tokensPadding;
// NB: Since 2023-10-14, the GPT-3.5 Turbo 0301 model shoves in 7 extra tokens to every message.
// More details: https://community.openai.com/t/gpt-3-5-turbo-0301-showing-different-behavior-suddenly/431326/14
if (queryModel.endsWith('-0301')) {
num_tokens += 7;
}
// not needed for cached tokenizers // not needed for cached tokenizers
//tokenizer.free(); //tokenizer.free();