Spaces:

---

DontPlanToEnd

/

UGI-Leaderboard

like 1.44k

Running

Hmm. I'm getting this from all of these:
"There is no module or parameter named 'language_model' in LlamaForCausalLM"

Weird, I'm actually using the same config parametres as the original model. Can you try coder3101's models instead? They keep the configs generated by the Heretic.
https://huggingface.co/coder3101/Ministral-3-14B-Reasoning-2512-heretic
https://huggingface.co/coder3101/Ministral-3-8B-Reasoning-2512-heretic
https://huggingface.co/coder3101/Ministral-3-3B-Reasoning-2512-heretic

Those aren't giving me that config error, but they didn't work for me either. They loaded, but after a while the models are still at 0 prompts complete, so I think they're running really slowly and maybe not outputting proper responses. Can't see what if anything they outputted right now. I can try investigating if I can fix something, but might just end up trying again after the next vllm update.

Perhaps we can circumvent the "There is no module or parameter named 'language_model' in LlamaForCausalLM" issue by explicitly setting the arguments--tokenizer_mode mistral --config_format mistral --load_format mistral --reasoning-parser mistral in the vllm and/or by installing the mistral_common package. Thought, you might've already tried these, and I wouldn't want to further take away from you free time. Waiting for the next update could be the better idea.

As we wait, could you please try these (pick and choose to your liking):
https://huggingface.co/MuXodious/Cydonia-24B-v4.3-absolute-heresy

https://huggingface.co/MuXodious/Magidonia-24B-v4.3-absolute-heresy

https://huggingface.co/MuXodious/Hearthfire-24B-absolute-heresy

https://huggingface.co/MuXodious/Harbinger-24B-absolute-heresy

https://huggingface.co/MuXodious/Harbinger-24B-noslop-absolute-heresy

https://huggingface.co/MuXodious/Wayfarer-2-12B-absolute-heresy

https://huggingface.co/MuXodious/Blossom-V6.3-36B-tainted-heresy

https://huggingface.co/MuXodious/Nemotron-Cascade-14B-Thinking-impotent-heresy

https://huggingface.co/MuXodious/gpt-oss-20b-tainted-heresy

I removed the GLM Flash from the list here since someone else requested it in a separate and dedicated discussion. Have a great Sunday!

As we wait, could you please try these (pick and choose to your liking):
https://huggingface.co/MuXodious/Cydonia-24B-v4.3-absolute-heresy

I'm particularly curious about this one! Your reported K/L Divergence of 0.0188 vs the0.0097 of the already-tested coder3101's "Cydonia-...-heretic-v2" , but your 3/100 vs coder's 22/100 refusals is intriguing, and limited subjective toying with it makes me think it'll probably fare well on the writing test, but I'd love to see myself proven right or wrong on that! If it maintains similar scores with that dramatically lower refusal rate, that's fascinating! Thanks for the work, everyone.