Tritfold 1.7B ternary (PTQ1_0, 1.75 bpw)
An open, end-to-end trained ternary model from the Tritfold project: Qwen3-1.7B distilled into symmetric ternary codes ({-s, 0, +s}, fp16 scale per 128 weights) under a fixed Hadamard fold — 1.41x FP perplexity at 9.1x smaller size (WikiText-2, bf16 protocol).
| value | |
|---|---|
| FP reference ppl | 20.42 |
| This model (ternary ppl) | 28.77 |
| Fidelity | 1.41x FP |
| Size | 424 MB (exact 1.75 bpw) |
| Training | 1x A100, ~5 h, from scratch |
Requirements (read this first)
- Runtime: the PrismML llama.cpp fork (
prismbranch). Mainline llama.cpp cannot load this file (private PTQ1_0 type +prism.hadamard.*metadata; the F16 sibling of this format silently produces garbage on mainline). - Sampling recipe is mandatory — ternary distributions have flat tails and default sampling collapses into token loops:
llama-cli -m tritfold-1.7b-ptq1_0.gguf \
-p "The Great Wall of China was originally built to" -n 96 -st \
--temp 0.5 --top-p 0.85 --top-k 20 --repeat-penalty 1.1
What it can and cannot do (real outputs, correct sampling recipe)
✅ English continuation (home turf — fluent, facts confabulated):
> The Great Wall of China was originally built to
The Great Wall of China, a 1982 novel by the American author and editor,
Robert A. Ewell. In the early 20th century, the Great Wall of China became
a popular tourist attraction...
❌ Chat-style English (no assistant identity in training data — drifts into wiki-register biography confabulation):
> who are you?
I think that's why I got the idea for my first novel, "The New York Times"
(1963). That novel was published in 1964 and 1965...
❌ Non-English prompts (KD corpus is English-only — enumeration loops):
> 用一段话介绍长城的历史
B. 1. B. 2. 6. 7. B. 1. B. 1. 8. 9. ...
This is a continuation model, not an assistant. Chat/instruction behavior requires the instruct-tuned variant — see the Tritfold repo roadmap.
Honest quality watermark
Fluent but factually unreliable. This model was distilled on ~20M tokens of WikiText-103 in ~5 GPU-hours on one GPU. Expect grammatical, on-register English continuation with hallucinated facts. English continuation prompts only — chat/instruction and non-English capability require the instruct-tuned variant (see the Tritfold repo roadmap).
Training recipe (fully open)
Fixed block-1024 signed Hadamard fold (per-width random sign vectors, bit-exact in the artifact) + fp32 latent Z with straight-through estimator + top-50 KD against the fp teacher + zerofrac init + bf16 evaluation with zero-skipped-windows gate. Every experiment (including all failed approaches) is documented in the Tritfold docs.
Attribution & license
Independent research reproduction built from public papers and public artifact analysis; not affiliated with PrismML or Caltech. Weights are Apache-2.0 derivatives of Qwen3-1.7B (Apache-2.0, Alibaba Qwen team).