Tritfold 1.7B ternary · instruct (PTQ1_0, 1.75 bpw)
The instruction-tuned variant of Tritfold: the wiki-only ternary model, further distilled on an ultrachat(60%)+wiki(40%) mix — it learns to answer, not to know.
| value | |
|---|---|
| Wiki ppl (bf16 protocol) | best 25.73 (1.26×FP) @step800 / 27.87 final (guard: never breached) |
| Size | 424 MB (exact 1.75 bpw, 9.1× smaller) |
| Instruction following (EN) | ✅ real behavior change — see before/after below |
| ARC-Challenge (likelihood) | 0.234 acc_norm (FP 0.377) — knowledge benchmarks are not fixed by chat distillation |
| Chinese | ❌ not acquired (EN-dominant corpus) |
Before / after (same prompt, same sampling, same runtime)
> who are you ? # wiki-only base model:
"I think that's why I got the idea for my first novel, 'The New York Times' (1963)..."
> who are you ? # this model:
"You're a person who is interested in science, and you're always eager to
learn more about the world around you. Here are some interesting facts..."
On-task, conversational register — though facts remain unreliable (the 1.75bpx watermark) and the identity answer is role-flipped (small-model quirk).
Requirements
- The PrismML llama.cpp fork (
prismbranch) — mainline cannot load PTQ1_0 or apply the Hadamard metadata. - Mandatory sampling recipe (ternary tails are flat; defaults loop):
llama-cli -m tritfold-1.7b-instruct-ptq1_0.gguf \
-p "List two tips for writing better Python code." -n 96 -st \
--temp 0.5 --top-p 0.85 --top-k 20 --repeat-penalty 1.1
- Cosmetic: outputs may start with a stray empty
<think>block (training-data artifact) — strip it if it bothers you.
License & attribution
Apache-2.0. Independent research reproduction from public papers and public artifact analysis; not affiliated with PrismML or Caltech. Weights are derivatives of Qwen3-1.7B (Apache-2.0, Alibaba Qwen team).