Tritfold 1.7B ternary · knowledge mix (PTQ1_0, 1.75 bpw) — v0.3
📓 Code, notebooks, all findings & full experiment log: github.com/benzeng/tritfold
Third variant in the Tritfold series: the instruct model further distilled on a knowledge-dense mix (FineWeb-Edu 3 + wiki 3 + ultrachat 2). Fixes the format, not the facts.
| value | |
|---|---|
| Wiki ppl (bf16 protocol) | best 25.20 (1.24×FP) — third consecutive record (28.77 → 25.73 → 25.20) |
| ARC-Challenge (likelihood) | 0.261 acc_norm — +11.5% over v0.2, first time above random (0.25); gate 0.265 missed by 0.004 |
| ARC generative (raw format) | 0.227 — format discipline fixed (unparsed 39% → 4.7%); letter knowledge remains coin-flip |
| Size | 424 MB (exact 1.75 bpw, 9.1× smaller) |
The knowledge ceiling (measured across 3 protocols)
| protocol | FP Qwen3-1.7B | this model |
|---|---|---|
| likelihood | 0.377 | 0.261 |
| generative, raw | 0.733 | 0.227 (parsed ≈ random) |
| generative, chat | 0.789 (parsed subset) | 0.244 (86% empty outputs) |
No protocol shows above-random letter knowledge at 1.75 bpw, while FP scores 0.73-0.79 in both generative protocols — the gap is capacity, not data recipe. ~12M tokens of educational distillation rebuild format and discourse, not facts.
> What planet is known as the Red Planet?
"The red planet is a star in the constellation of ... one of the most famous
stars in the universe" # discourse intact, facts confabulated (no Mars)
Requirements
- The PrismML llama.cpp fork (
prismbranch) — mainline cannot load PTQ1_0. - Mandatory sampling recipe:
--temp 0.5 --top-p 0.85 --top-k 20 --repeat-penalty 1.1
License & attribution
Apache-2.0. Independent research reproduction; not affiliated with PrismML or Caltech. Derivative of Qwen3-1.7B (Apache-2.0, Alibaba Qwen team).