Disclosure: I build Grunz, which runs abliterated open models, so this is from the using side rather than from having abliterated Krikri myself. One pitfall that isn’t in the thread yet: test instruction following and output format, not just refusals. What we keep seeing with abliterated variants is that knowledge and fluency survive fine, but multi-step instructions, sticking to a requested format, and stop behaviour degrade first. A refusal check will say it worked and a quick chat will feel fine, then it falls apart the first time you ask for structured output. An 8B has less margin there to begin with. Cheap way to catch it: write ten or so Greek prompts with strict format requirements (a JSON object, a numbered list of exactly five items, answer in one sentence) and score base vs abliterated side by side before you quantize, then again after Q4_K_M. We suspect quantization damage compounds with abliteration damage, so if Q4 looks shaky on those, Q5_K_M or Q6_K is worth it, and with 117 GB you can easily afford that for an 8B. On workflow I’d abliterate the safetensors and convert to GGUF afterwards, so the direction is computed at full precision.

Hey everyone, Setup: GMKtec EVO-X3, AMD Ryzen AI MAX+ 395 (Strix Halo), Radeon 8060S, 117.55 GiB unified memory, Ubuntu 24.04, Vulkan/RADV, llama.cpp built with GGML_VULKAN=ON. Running Qwen3-32B-heretic-Q8_0 on llama-server (127.0.0.1:8081). Goal: build an uncensored/abliterated Greek-native LLM for local use. Best candidate: ilsp/Llama-Krikri-8B-Instruct (Llama 3.1 8B Greek fine-tune). Plan: abliterate with Heretic (heretic-llm), convert to GGUF Q4_K_M, run on Vulkan. Questions: 1. Has anyone abliterated Llama-Krikri-8B or any Greek fine-tune? Does Heretic support it? 2. Best workflow: abliterate safetensors then convert to GGUF, or use abliterate.cpp on GGUF? Any pitfalls? 3. Best quantization for Greek on Strix Halo: Q4_K_M, Q5_K_M, Q6_K? Need quality/speed balance. 4. Recommended llama.cpp Vulkan flags for Strix Halo: -ngl 99, -c 32768, flash attention, q8_0 KV cache, speculative decoding? Any known issues? 5. Any existing uncensored Greek models I missed (Krikri, Meltemi, Sophea, etc.)? Thanks!

Πώς η γαλακτοβιομηχανία εξελίχθηκε σε ηγέτη της κατηγορίας του γιαουρτιού, η επένδυση για κατασκευή νέων μονάδων παραγωγής και τα σχέδια για είσοδο σε ΗΠΑ και Κίνα.

Πώς η γαλακτοβιομηχανία εξελίχθηκε σε ηγέτη της κατηγορίας του γιαουρτιού, η επένδυση για κατασκευή νέων μονάδων παραγωγής και τα σχέδια για είσοδο σε ΗΠΑ και Κίνα.