JanV1-Q8 still cant answer some basic of questions
Posted by choxxolatee@reddit | LocalLLaMA | View on Reddit | 18 comments
From a post 3 months ago (link), OP showed how broken JanV1 was. Emre from Jan replied and suggested using Q8 with adjusted parameters and Serper tool, Emre then attached a few screenshots in which they ran the exact same question as OP and gave a correct answer.
I tried to replicate it today with the same Model, parameters and questions and I was given the wrong answer. I asked the same question about the GDP of US
I then asked about the stock price of Nvidia
egomarker@reddit
Show the result of a tool call.
choxxolatee@reddit (OP)
https://imgur.com/a/N1BZfm4
egomarker@reddit
choxxolatee@reddit (OP)
Is there anything im doing wrong to not get these results??
egomarker@reddit
My only guess is your search tool returns something weird.
choxxolatee@reddit (OP)
May you please help me? I tried using your model (Janv1 2509) and search tool (tavily) and I am still getting bullshit results?
egomarker@reddit
Upload the full json returned by search somewhere e.g. privatebin.net and give me the link.
choxxolatee@reddit (OP)
https://privatebin.net/?6bbb6fd22abb8d1c#Ab1NZ6SWQRd3XfAMV3p84vJjMRDzFRLbxZmxnaWFq2xj
egomarker@reddit
Kind of works even with no tool calls at all on my side..
Maybe your system prompt is off?
choxxolatee@reddit (OP)
My prompt is the exact same as yours and its just a very basic question. Also, I turned off all MCP and just tried with no tools. Is my AI broken??
egomarker@reddit
Maybe try changing your assistant to default Jan
choxxolatee@reddit (OP)
I doubt it because I barely change the wording for the assistant. Im complete noob in coding and would like to ask whether I skipped some steps? Do I need to install dependencies to run Tavily servers? All I did was create a new MCP server within Jan and pasted the API key. Linked here https://privatebin.net/?97e89160fa820e3f#3rXmgGahRmo5mbuubQY29Z9mQjTz5VLGQz6UrtFGBazc
egomarker@reddit
It doesn't matter, model gives me 30 trillion GDP even without web search at all in empty chat.
choxxolatee@reddit (OP)
may I ask what app are u using? are you using JanAi or other local LLM ? why does ur format look different to mine?
egomarker@reddit
LM Studio
Wise-Carry6135@reddit
yeah look at the tavily response - if bad use another one like linkup.so
choxxolatee@reddit (OP)
I used several tools at this point. None worked. Brave serach/ Browser MCP/ Serper and now Tavily
egomarker@reddit