neuroscr/computer-10-GGUF

computer-10 (GGUF)

GGUF builds of cosmicoptima/computer-10 — a character model, not a helpful assistant.

What it is

Computer-10 is trained from Llama 3.1 70B base with no assistant data. The point is a distinct personality (“Model C”), not chat-tuned helpfulness:

• unusual, opinionated voice
• expresses views freely
• not designed to be servile or “happy to help”
• closer to a character than to Instruct / ChatML assistants

If you prompt it like ChatGPT, you will be disappointed. If you talk to it like a weird interlocutor in its native format, that is the intended vibe. Author context: @parafactual.

Training (from the upstream card)

This checkpoint is Computer-9 further trained on its own conversations (self-distillation):

• 25,197 Model C turns that passed a mechanical filter and a logic check
• length- and surprisal-matched across bins
• only Model C’s turns in the loss (user side not trained as assistant)
• full-parameter, one epoch, lr 2e-6
• nothing in the text was hand-edited

So the “assistant” stack of refusals / corporate tone from Instruct never got baked in; the model stays in the Computer lineage’s document-style dialogue.

Files

File	Type	Size (approx)	Notes
computer-10-bf16.gguf	BF16	~132 GB	lossless vs upstream bf16 safetensors
computer-10-Q4_K_M.gguf	Q4_K_M	~40 GB	good default for 2×24 GB GPUs

Converted with llama.cpp (convert_hf_to_gguf.py --outtype bf16, then llama-quantize … Q4_K_M).

Prompt format

No chat template. Same as the other Computers:

Full conversation with Model C:
**User:** …
**Model C:** …
**User:** …
**Model C:**

Suggested sampling (upstream): temperature 1.0, top-p 0.98, stop on \n\n**User:**.

Intended use / non-use

• Use: character dialogue, open-ended conversation, exploring a non-assistant personality.
• Don’t expect: RLHF-style deference, tool-calling Instruct habits, or “as an AI language model…” framing.
• License: Llama 3.1.

Source

Upstream weights and training note: cosmicoptima/computer-10.