This model makes no sense

#3
by rombodawg - opened

Is this a finetuned model? If so why is it catagorized as a quantization of (DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking).

Did you just rename their model after quantizing it and not change anything? Because all that does is confuse people as to the purpose of your model.

Please clarify this issue

BugTraceAI org

Hey rombodawg β€” fair to raise the question, so let me be transparent about what actually happened here.

This is a fine-tune, not a rename. The "quantization of" tag in the HuggingFace metadata is auto-generated based on the base_model field β€” it doesn't capture the full lineage. Here's the actual pipeline:

Base selection β€” We chose DavidAU/Qwen3.6-27B-Heretic2-Uncensored-Finetune-Thinking specifically because the community was asking for Qwen3 architecture (see discussion #1). Starting from an already-abliterated base means we're not wasting compute re-teaching it to be uncensored.

Custom dataset β€” 2,541 training examples curated from: HackerOne disclosed reports, CVE writeups, Bug Bounty Daily CoT chains, GitHub security research (2024–2026). Not publicly available as a single dataset anywhere. This took weeks of ETL pipelines, scraping, normalization, and DPO-format conversion (including a β†’ tag migration for Qwen3's native vocab).

Training β€” SFT via Unsloth on a RunPod H100 80GB, LoRA r=16 injected across all 7 attention + MLP modules across 64 transformer layers, 2 epochs. The VRAM math alone required significant engineering work (54GB base model + optimizer states).

Merge + IMatrix Quant β€” LoRA merged back to BF16 full weights, then quantized with IMatrix guidance via llama.cpp to Q4_K_S and Q6_K.

Benchmarked β€” 5/5 on our BugTraceAI Ultra Bench v1.0 (Nuclei templates, CVE PoCs, JWT cracking, kernel exploits) at 0% refusal rate.

All of this is already documented in the model card under 🧠 Training Details β€” worth a read before assuming. The base model is the foundation, not the product. The work is in the data, the training objective, and what the model now does that the base didn't.

Can we possibly have the q8 quantized version as well? q6 seems to be hallucinating and jumping to conclusions. It falsely identified cloakbrowser as fully open source when it's only half open source and the actual browser build is closed source proprietary . Overall it seems weaker than the q8 qwen 27B I've been using. It degraded almost to ornith 35B MoE level where it had only shallow reasoning and didnt look deep into things

(Btw I'm sure nobody wants to hear this, but as far as "consumer grade GPUs" are concerned, an a6000 with 48GB of VRAM now counts as high end consumer GPU, a 4090 with 24GB of VRAM is considered entry level, and anything below that... well.... it's not pretty. What is a single rtx 6000 pro considered? Something like a gaming rig or stuff for creators to use. Now, only MULTIPLE rtx 6000 pros are considered "out there". I say this because I learned the hard way that the minimum quality a gguf can have is q8 to retain some semblance of the awesomeness of the full weights... )

Sign up or log in to comment