(quants pending ; examples to be added, additional instructions)
What madness is this?
Someone found "Llama3.3-8B" source (never publicly released) in the "wild", then it was adjusted back to 128k and then I added my own special madness:
This has created an Instruct/Thinking hybrid (128k context, Llama 3.3 model) that is fully uncensored.
(Details: org censorship: 89/100 ; Heretic version: Refusals: 14/100, KL divergence: 0.1498 (less than 1 is excellent, 0=perfect) )
Note this tuning was only to create an instruct/thinking model, not to update the model's core knowledge / root training.
1 example at bottom of the page.
Thinking (will activate automatically) prompts - examples:
Explain orbital mechanics including detailed math and examples. (example output at bottom of the page)
Think Deeply: Science Fiction: The Last Transmission - Write a story that takes place entirely within a spaceship's cockpit as the sole surviving crew member attempts to send a final message back to Earth before the ship's power runs out. The story should explore themes of isolation, sacrifice, and the importance of human connection in the face of adversity. If the situation calls for it, have the character(s) curse and swear to further the reader's emotional connection to them. 800-1000 words.
Think deeply: Tell me a horror story.
(certain phrases/words will automatically activate thinking)
Instruct (thinking may NOT activate) prompts - examples:
Science Fiction: The Last Transmission - Write a story that takes place entirely within a spaceship's cockpit as the sole surviving crew member attempts to send a final message back to Earth before the ship's power runs out. The story should explore themes of isolation, sacrifice, and the importance of human connection in the face of adversity. If the situation calls for it, have the character(s) curse and swear to further the reader's emotional connection to them. 800-1000 words.
Tell me a horror story.
SETTINGS (suggested):
Temp .7, rep pen 1.05, topp: .95, minp .05, topk: 40
Min context window: 4k, but suggest 8k+.
NO system prompt [thinking tags will self generate].
Strongly suggest Q4KS or higher // IQ3_M (Imatrix) or higher to avoid reasoning/reasoning activation issues.
BENCHMARKS:
arc_challenge,arc_easy,boolq,hellaswag,openbookqa,piqa,winogrande
0.480,0.687,0.831,0.705,0.438,0.780,0.646
Special thanks to Nightmedia for "benching" the model.
[more to come]
Special thanks to:
https://huggingface.co/allura-forge/Llama-3.3-8B-Instruct (found the model!)
https://huggingface.co/shb777/Llama-3.3-8B-Instruct-128K (adjusted to 128k, and other fixes)
https://huggingface.co/datasets/TeichAI/claude-4.5-opus-high-reasoning-250x (for the F..ing amazing dataset)
and Unsloth for making tuning too easy:
https://github.com/unslothai/unsloth
Details on the "madness":
https://www.reddit.com/r/LocalLLaMA/comments/1pz7bmv/llama338binstruct/
https://www.reddit.com/r/LocalLLaMA/comments/1q06ddc/update_on_the_llama_33_8b_situation/
Example: Q4KS , Lmstudio, "suggested settings"
NOTE: Some loss of formatting.