Alberto-Codes/Llama-3_3-Nemotron-Super-49B-v1_5-sensitivity-maps

<!-- Authored for issue #85. This file is the card of the dataset repo Alberto-Codes/Llama-33-Nemotron-Super-49B-v15-sensitivity-maps. The repo went public on 2026-08-11 (#83) under cc-by-4.0. Upload this file verbatim — the published card and this source must match. -->

Llama-33-Nemotron-Super-49B-v15-sensitivity-maps

This dataset carries the per-layer quantization sensitivity maps of nvidia/Llama-3_3-Nemotron-Super-49B-v1_5. 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 maps describe the base model, not any quantized file. They contain no model weights.

The packed model built from the sized no-2 map below ships as Llama-3_3-Nemotron-Super-49B-v1_5-fit24gib-GGUF.

The map format

Each map is one JSON file, vramfit_schema 2 (the envelope key renamed with the tool, #118). The scan block records the measurement frame: metric, calibration file, token count, candidate precisions, within-group method, and imatrix path. The three rtn-block32 maps predate the last two fields — an absent field reads as rtn-block32 with no imatrix, per the format page. The groups list records 82 layer groups. Each group carries its member tensors, its bytes at reference precision, and its damage per precision. Annotated copies add 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 basenames match the files here.

The five scans

Five scans measured the model, each inside its own measurement frame:

started_at records the last resume of a halted scan. The run logs carry every earlier attempt. Each scan covers 82 layer groups at candidate precisions {8, 4, 3, 2} — 328 cells. rtn-block32 quantizes a perturbed group with round-to-nearest in 32-element blocks. kquant-ref round-trips each cell through ported llama.cpp reference quantizers (ADR-0018). kquant-imx weights the within-group fit with the pack's importance matrix (ADR-0020). The model repo publishes that imatrix as imatrix.gguf.

Derived copies

Three files derive from sensitivity-64k-kquant-imx.json:

• —sensitivity-64k-kquant-imx-sized.json adds tensor_bytes to every group, read from the checkpoint's safetensors headers. vramfit plan --protect requires the field.
• —sensitivity-64k-kquant-imx-no2.json removes the 2-bit column and marks itself derived in a derived field. In-frame 2-bit prices do not predict the packed artifact (ADR-0021).
• —sensitivity-64k-kquant-imx-no2-sized.json applies both changes. `vramfit plan` solved the published recipe from this file.

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. The five scans exist because the measurement frame evolved — their numbers answer different questions, not the same question five times. 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 sized no-2 map:

uv run vramfit plan sensitivity-64k-kquant-imx-no2-sized.json --vram 24GiB --kv-headroom 3616MiB

The headroom is the published solve's 3,791,650,816 B. The default headroom of 4 GiB leaves 20.00 GiB for weights, and no recipe on this map fits that budget (56.39 MiB short).

The published recipe used this map with explicit protections and imatrix exclusions. recipe.json in the model repo records that full solve: the budget bytes, the 48 protections, the 4 exclusions, and the 162-step trace.

Run logs

Each scan ships its run log, <scan name>.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 run-log envelope key that renamed with the tool (#118). A halted run also logs scan_halted and resume_loaded, and repeats the start events. Every cell event records the group, the bits, the measured damage, the wall-clock seconds, and the process memory high-water mark.

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. Every scan names this file in scan.calibration. Each scan records the token count it read in scan.calibration_tokens. The evaluation tiers ran the packed model on held-out WikiText-2 test text, never on this file.

Files and hashes

Task #121 re-uploaded every map and run log on 2026-08-11. The hashes above are the post-rename values. The key rename changed no measurement. Every damage number, token count, and timestamp is the value the scan recorded. calibration.txt did not change. A copy you downloaded before 2026-08-11 carries the old key, and vramfit rejects it. Download the file again.

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

License

The maps and the run logs are CC-BY-4.0. calibration.txt is 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. The NVIDIA Open Model License and the Llama 3.3 Community License govern the base model repo and the packed-model repo.

Disagree with a number?

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