Baekpica/Inkling-Small-Multimodal-Calibration

Inkling-Small Multimodal Calibration

The exact 1,663 samples used for BF16 routed-expert importance collection for Inkling-Small Mixed Quant GGUF. This is calibration material, not a held-out evaluation benchmark.

The primary balanced pass is:

A subsequent coverage pass added 201 real structured reasoning/tool conversations (529,126 tokens, seed 3408, up to 4,096 tokens per sample). It probes rare experts, including L41 expert 123, which remained unobserved after both passes. The final model protects L40–41 with Q4K; expert 123 uses ordinary Q4K without fabricated importance data. These are normalized from the pinned NVIDIA Instruction-Following-Chat-v2 reasoningon and Cascade-SFT-Stage-2 toolcalling sources and rendered with actual Inkling thinking/tool syntax. Total actual input: 1,578,255 tokens.

importance_weight is 1.0 for primary examples and 0.1 for the coverage supplement. Imatrix statistics normalize weighted squared-input sums by weighted actual hit counts; raw unweighted route counts are audited separately. The weights keep rare-expert coverage additions from dominating the original multimodal mix.

The original balanced dataset remains available at revision 4ceb58745dcb2cde3d018c52942f881f072284ac. This revision includes both passes. The effective weighted token count is 1,102,041.6; it is not a count of unique observations. The public manifest records both counts explicitly.

Shares count valid decoder positions, including image/audio placeholder positions and their accompanying text. They are not document shares, media-only token shares, or separately reweighted loss quotas. Padding does not contribute. Each sparse layer routes six expert observations per valid decoder token.

Text is sampled deterministically (seed 3407) from checksum-verified prior K2-Horizon calibration text pools, using the composition of the Solar Healing Mix and a Tulu/FineWeb supplement. These pools were decoded with the Solar tokenizer; we re-tokenized the text using Inkling, without carrying Solar token IDs or its control-token syntax into the model. Fixed 1,024-token text windows may cross original document boundaries. Code/tool denotes the source category, not a claim that every sample is a complete executable tool conversation.

Image inputs use the first training parquet of DocVQA and ChartQA at pinned revisions; duplicate image bytes are rejected. Images are resized within 1280×1280 before the official Inkling processor. Audio uses real LibriSpeech train.clean.100 recordings at 16 kHz (2–60 seconds), the official dMel processor and corresponding transcription text. No synthetic media or random embeddings stand in for a modality. source records identify the original parquet row.

The Parquet index contains readable text, input IDs, category and exact tensor pack filename/hash. The actual processor outputs are in tensors/*.safetensors: input_ids, optional attention_mask, pixel_values, audio_input_ids, and audio_input_ids_mask. These allow bit-exact input replay without decoding the original media again. They are model inputs, not learned model weights.

from datasets import load_dataset
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
rows = load_dataset("Baekpica/Inkling-Small-Multimodal-Calibration", split="train")
row = rows[0]
path = hf_hub_download("Baekpica/Inkling-Small-Multimodal-Calibration",
                       row["tensor_file"], repo_type="dataset")
inputs = load_file(path)

The tokenizer, source configuration, chat template, processor configuration, selection manifest and source lineage are included. Input processing used Hugging Face Transformers source checked out on 2026-09-09; the exact commit is recorded with the reproduction materials and model handoff.

See LICENSES.md. Source-specific licenses remain applicable; this collection does not relicense its constituent examples.

Rebuilding and replay

For exact replay, download the published tensor packs and use the index above. build_samples.py reconstructs the primary pack from source media and the checksum-pinned K2 text pools; those historical pool files are retained in the owner’s earlier private handoff. build_supplement.py reconstructs the added structured dialogs; its two source dataset revisions and raw row hashes are in manifest.json. The normalizer helpers are under refs/. Set INKLING_ROOT to the prepared work tree, with tokenizer/processor files in source/, source media under calibration/, and the helpers in refs/. The pinned Transformers commit is in provenance/toolchain.json. Exact input replay does not require access to the private historical text pools.