NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF-metrics

# NVIDIA Nemotron 3.5 Bolt E4B GGUF: metrics Everything behind the numbers in [AtomicChat/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF](https://huggingface.co/AtomicChat/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF): the importance matrix, the per-tensor type rules of every file, and the quantize and KL divergence logs. Stand: BF16 reference converted with upstream llama.cpp 36e36e63; KL divergence and top-1 agreement against it on 24 chunks of 4096 tokens of `eval_neutral` (30 languages, never used for calibration) from [AtomicChat/calib-corpora](https://huggingface.co/datasets/AtomicChat/calib-corpora); every file, ours and other publishers', run with the same reference, text and build on one A100 40GB. | Folder | What is in it | | --- | --- | | `imatrix/imatrix.gguf` | Our importance matrix: BF16 weights, 9,377 chunks of 512 tokens of the `nemotron-bolt-e4b` corpus (rendered through this model's chat template), `--parse-special` | | `fast/` | The first release-day build (layout v1, Unsloth's published imatrix): ladder rules, quantize logs, `kld.md` with our files and Unsloth's, LM Studio's and NVIDIA's on the same stand | | `v2/` | Layout experiments at the Q4 point (head or per-layer table back at q8_0, the whole FFN one step up) and the v2 ladder | | `v3/` | The published layout (attention one or two steps up): ladder rules, quantize and KLD logs with Unsloth's and our imatrix, and the held-out code check | Which imatrix each published file uses: `AD-Q4_K_M` ours; `AD-Q5_K_M`, `AD-IQ4_XS`, `AD-IQ3_S` Unsloth's published one; `AD-Q6_K` (layout v1) Unsloth's as well; `Q8_0` uses none. The type of every tensor of a file is in the `.types` file of its rung under `ladder/` (the first rule that matches a tensor wins).