Qwen3-1.7B GPTQ Hessians — wikitext-2 train ONLY
⚠️ READ THE CORPUS LINE BEFORE REUSING THESE.
| model | Qwen/Qwen3-1.7B |
| corpus | wikitext-2-raw-v1 / TRAIN — single domain, Wikipedia prose |
| budget | 1024 rows x 256 tokens = 262,144 tokens |
| geometry | g128, the 67-geometry (--rotate auto --rotate-cap 128 --rot-pairs down_proj=256) |
| storage | UNROTATED — one map serves any --rotate width; R^T H R is applied after loading |
| files | gptq-rot-g128-1024x256/block000..027.safetensors, 28 blocks, 7.05 GB |
What these are valid for
H = sum x^T x is built from the QUANTIZED PREFIX: block k's Hessian
comes from activations produced by blocks 0..k-1 after they were
quantized. So these are valid only for placing a body from the fp16
base. A run whose embedding (or anything upstream) is already ternary sees
different activations and must build its own.
What they are NOT
They are not an ULTRAMEGA Hessian. They saw Wikipedia prose and nothing else — no code, no math, no chat, no reasoning traces. A placement fitted to them is tuned to one distribution, and graded on wikitext-2 test it will look better than it deserves on anything wider. For work that must generalise, build H on the multi-lane calibration corpus instead.
Usage: ternary_gptq.py --load-hessians <dir with gptq-rot-g128-1024x256/>