gemma-4-31B-it-sensitivity-maps

This dataset carries the per-layer quantization sensitivity map and the in-frame importance matrix of google/gemma-4-31B-it, measured on the QAT unquantized checkpoint at revision 1e4d8beecacb8b7590c1d8bedd7335f687bf311f. vramfit measured them. A sensitivity map records one damage number per layer group and candidate precision. Damage is the shift in the model's output distribution when that group alone quantizes — mean final-logits KL divergence against the bf16 reference. The map describes the base model, not any quantized file. It contains no model weights.

The packed model solved from this map ships as gemma-4-31B-it-fit24gib-GGUF. Its card states that the map and the matrix stay in the run archive. Issue #449 ruled on 2026-09-04 that both publish here.

The map format

The map is one JSON file, vramfit_schema 3, in the form that carries group_by: layer. The scan block records the measurement frame: metric, calibration file, token count, candidate precisions, grouping, within-group method, and imatrix path. The groups list records 61 decoder groups: the token embedding and the 60 language-model layers. Each group carries its member tensors, its bytes at reference precision, its damage per precision, and tensor_bytes, the per-tensor size split. The sensitivity map format page specifies every field. The path fields record the reference box's absolute paths. The scan.calibration and scan.imatrix basenames match the files here.

The map holds the language model only. The vision tower and the projector never enter it: the published pack ships the vendor's projector as a separate sidecar, converted to Q4_K_M without a measurement.

The scan

File	Calibration tokens	Within-group method	Imatrix	Started (UTC)	Cells
sensitivity-32k-kquant-imx.json	32,768	kquant-imx	yes	2026-08-27	244

The scan covers the 61 groups at candidate precisions {8, 4, 3, 2}, which is 244 cells. The kquant-imx method round-trips each cell through llama.cpp's k-quant types with the importance matrix below. The scan ran on the reference box under a 12 GiB GPU memory cap with the remaining groups offloaded to host memory, and finished in one invocation (run id d4e07748ae2a, 2026-08-27 02:46 to 16:06 UTC).

The published recipe solved from this map. recipe.json in the model repo records the full solve: the 24 GiB budget, the 9 GiB KV headroom, the 0.005 format overhead, and the 81-step trace.

Do not compare damage across files

Damage values are calibration-relative and frame-relative. They compare only within one file. Do not rank damage across scans, across calibration sets, or across models. Rank packed models by measured quality at a fixed model and budget, never by raw damage.

Solve a recipe

vramfit plan is pure Python and imports no torch. Solve your own budget against the map:

uv run vramfit plan sensitivity-32k-kquant-imx.json \
  --vram 24GiB --kv-headroom 9GiB --format-overhead 0.005

Those three values reproduce the published solve. Pass your own --vram and --kv-headroom for a different budget. recipe.json in the model repo records every resolved value.

The importance matrix

gemma-4-31b-bf16-framed.imatrix.gguf is the matrix the scan and the pack consumed. llama-imatrix b10362 built it with --parse-special over calibration-framed.txt: 356 chunks at n_ctx 512 through the BF16 decoder GGUF converted from the same checkpoint at the same revision. The pack step consumed this file by name, so a pack that passes it reproduces the published bytes.

Coverage derives from the matrix's own entry names, never from a label. Two absences matter:

• The matrix carries no token_embd.weight entry, so the token embedding quantized unassisted. This is expected at b10362.
• The matrix carries attn_v entries for 50 of 60 layers. The 10 full_attention layers (5, 11, 17, 23, 29, 35, 41, 47, 53, 59) receive no attn_v activation from the b10362 graph, so those ten tensors quantized unassisted. This is a property of the instrument, not a defect in the matrix.

Run log

The scan ships its run log, sensitivity-32k-kquant-imx.runlog.jsonl — structured JSONL, one cell_measured event per cell between the lifecycle events scan_started, meter_built, and scan_finished. Every line carries vramfit_runlog 2. The scan ran start to finish with no halt: 244 cell events. Every cell event records the group, the bits, the measured damage, the wall-clock seconds, and the process memory high-water mark. The meter_built event lists the vision-tower tensors the matrix does not cover. The scan never measures them.

The calibration set

calibration.txt is the complete Project Gutenberg ebook of Pride and Prejudice, unmodified, with the Project Gutenberg header and license text intact — byte-identical to the calibration file of the project's two earlier map datasets.

calibration-framed.txt is the same text inside the checkpoint's own answer channel: 357 blocks of about 512 tokens, each wrapped in the Gemma chat template by vramfit's frame_calibration.py. Gemma 4 31B IT-QAT prices raw prose at a perplexity near 3,000 and the same prose inside its own channel at 26 to 75, so every measurement here ran on the framed file. The scan names it in scan.calibration and reads 32,768 tokens of it. The matrix ran over all 182,404 tokens. The evaluation tiers ran the packed model on held-out WikiText-2 test text, never on either file.

Files and hashes

File	SHA-256
sensitivity-32k-kquant-imx.json	553e13ab9e3f1b34291beed8acf53c14e04df20e2383921e4b4ae6cad4d931d7
sensitivity-32k-kquant-imx.runlog.jsonl	7461f0cdbaa72b003f8b4017ced61db85b3070ef1d6aa1dc611c43dd6ca23e22
gemma-4-31b-bf16-framed.imatrix.gguf	4a168bd7309f787d89f237ff84520980a4b9995975a31c758fd69e8e9d275c3d
calibration.txt	74f2665d6e6925fc2c17dec644bec9e87df478a0f1836822125e8acbb3777806
calibration-framed.txt	98ab7220cdda1c1c6cd57ccf072daa2a1d0a5890ad82f25e982d5bbc8234c55d

Hashes prove identity, not quality. The measurement evidence is the run log beside the map.

License

The map and the run log are CC-BY-4.0. The importance matrix is activation statistics of the Gemma 4 checkpoint, which Google DeepMind released under the Apache 2.0 license — see the Gemma 4 license note. The two calibration files are a Project Gutenberg ebook, public domain in the United States, distributed with its Project Gutenberg header intact. This dataset carries measurements of the base model, not the base model's weights.

Disagree with a number?

Re-run the scan. The vramfit repository documents the scan command, the meter, and the settings the run log records. A map you measure yourself beats one you argue with.