About the Qwen 3.5 9B Heretic (Sorry for topic unrelated to this model)
Hello darkc0de,
I've been using the Qwen 3.5 9B Heretic that you created a while ago when Qwen 3.5 came out. I've tried multiple different Heretic versions of this model, but yours was probably the best version. Ironically, it is gone now. Where did it go? You make some of the best finetunes, but they don't seem to remain available that long. 😕
And again, sorry for off-topic, I didn't know where else to ask this.
A couple things....
1.) Huggingface began charging me a bunch of $ for storage. With HF Pro I get 1TB private storage, 11.2TB Public storage. I was way over my limits and didn't notice until they suspended my account for non-payment.
2.) I realized the vast majority of my profile was basically experiential garbage, much of which didn't even work, or models so old they were basically obsolete at this point.
So I began to cleanup my profile which huggingface doesn't necessarily make it easy to bulk delete repos, and I had hundreds of repos to get rid of so... I proceeded to bitch whine and complain like a baby to huggingface... Thus they sent me some vibe coded python script to bulk delete some of my repos, which malfunctioned, which pissed me off further... So naturally I spent a whole day n a half rage deleting all my repos manually one by one, while cursing at my monitor, chugging code red, and chain smoking newports 🚬
The Qwen3.5-9B tune was likely a victim said rage. 🤬
I only really made sure to keep my most popular couple models and a few very current projects. I just said fk it and let it all burn. Maybe someone in the community has a copy or maybe it was quant'd by someone somewhere I don't know.
A couple things....
1.) Huggingface began charging me a bunch of $ for storage. With HF Pro I get 1TB private storage, 11.2TB Public storage. I was way over my limits and didn't notice until they suspended my account for non-payment.
2.) I realized the vast majority of my profile was basically experiential garbage, much of which didn't even work, or models so old they were basically obsolete at this point.So I began to cleanup my profile which huggingface doesn't necessarily make it easy to bulk delete repos, and I had hundreds of repos to get rid of so... I proceeded to bitch whine and complain like a baby to huggingface... Thus they sent me some vibe coded python script to bulk delete some of my repos, which malfunctioned, which pissed me off further... So naturally I spent a whole day n a half rage deleting all my repos manually one by one, while cursing at my monitor, chugging code red, and chain smoking newports 🚬
The Qwen3.5-9B tune was likely a victim said rage. 🤬
I only really made sure to keep my most popular couple models and a few very current projects. I just said fk it and let it all burn. Maybe someone in the community has a copy or maybe it was quant'd by someone somewhere I don't know.
That's sad, thanks for explanation and I'm sorry to hear that you had to go through so much trouble here.
Anyway, the model was quantized into GGUF by mradermacher, so at least it is preserved in that format for now and it is what I'm currently using. Unfortunately, if anyone wanted to use that model for some merges or further finetuning, that's not gonna work with GGUF format. The original format in safetensors would be needed.
GGUF quants are available here: https://huggingface.co/mradermacher/Qwen3.5-9B-heretic-GGUF
Damn, sorry this happened. I guess that's why your Hermes Heretic ain't there anymore either, do you happen to remember the KL div and refusal rate for that one (so I can copy paste the info in my GGUF repo)?
Anyone bothered to make a clone at least? I really need one. I was about to start a lora last night when I saw the 404 and was like daaaaamn. But it's ok.
I can get another heretic copy likely from Fave and do the job. Hope you're good. I keep an eye on you.
Mi spiace verramente tanto. A presto
I'm currently working on a new tune of this 27B, if everything works out it'll probably be ready to release in about a week or so.
Currently generating the dataset on a NVIDIA RTX PRO 6000 Blackwell Workstation Edition 96GB
Still has about 32 hours left to generate, then I'll start training probably for a single epoch.
If it works well I'll repeat the process for 9B
cool... any news on the 9b? or we keep an eye on you?
Buona pasqua! <3