Danube3-4b-4chan-HESOYAM-2510

Model Details

I finetuned Danube3 4B Base on adamo1139/uninstruct-v1-experimental-chatml dataset with the goal being making AI assistant slop less likely.

Then I did finetuning on adamo1139/4chan_archive_ShareGPT_only5 which is a filtered collection of 4chan threads from various boards for 1 epoch to introduce 4chan-specific slang.

Then I did finetuning on adamo1139/HESOYAM_v0.4 for 3 epochs to improve 1-on-1 chat capabilities.

This is a resulting model.

Prompt format

Use ChatML prompt format.

System message should be in the format as below:

A chat on 4chan board /3/
A chat on 4chan board /g/
A chat on 4chan board /x/
A chat on 4chan board /pol/

Evaluation

I am still vibe-checking the model but initial results are good. I might have put in a bit too much reddit style from HESOYAM, not sure.

GGUF Quants

Quants are available here: adamo1139/danube3-4b-4chan-hesoyam-2510-gguf

Training details

Training details and LoRA adapters can be made available if you request it. I am just not sure if anyone is interested in them, so I am not uploading those artifacts yet.

Future plans

A bit of experimentation on the 4B/500M Danube3 models and then I want to improve my current Yi 34B 200K HESOYAM 0208 model using 4chan archive dataset.