QLNI/shadow-50m-vision

SHADOW 50M VISION: proof run

A 50M-class ternary (1.58-bit) model that reads pictures, clips and sounds as words in one frozen 512-bit table shared with language, keeps its knowledge in editable on-disk Engram tables, and runs on a laptop CPU. This repository is the proof run in one place: the complete training data as streamable token ids, the table that gives the ids their meaning, the scripts, and (after training) the checkpoint and results. Nothing else has to be downloaded to reproduce it.

Non-commercial research release.

Stream it

from datasets import load_dataset
ds = load_dataset("QLNI/shadow-50m-vision", "understanding", split="train", streaming=True)
doc = next(iter(ds))        # {"ids": [...], "source": "pixmo-cap", "licence": "odc-by"}

One row is one document: ids are rows of the frozen table (table/table.npz), mask (SFT only) marks the tokens trained on (the model's turns), source and licence say where it came from. One source alone: data_files="data/understanding/pixmo-cap/*.parquet".

The data: 2.182B tokens, 40 sources

lane	tokens	sources
knowledge	1,158M	cosmopedia-v2 college and high-school textbooks 349M, Ultra-FineWeb-L3 289M, Ultra-FineWeb 175M, OpenMathInstruct-2 138M (GSM8K-style 97M, MATH-style 42M), OpenCodeInstruct 81M, Wikipedia 59M, FineMath-4+ 35M, UltraData-Code 32M
understanding	576M	FineVision 187M, OBELICS 163M (222k web pages, their pictures inline), PixMo-cap 136M (584k dense captions), AudioSkills-XL 20M, AudioSet 6M, WavCaps 4M, MSR-VTT 2M, LLaVA-Video 3M, Clotho 1M; generation pairs in both orders (caption [gen_image] words [end_gen] and the reverse): FLUX-Reason-6M 32M, OpenVid-1M 9M, AudioCaps 7M, WavCaps 4M, Clotho 2M, MSR-VTT 1M
SFT	447M	Nemotron-Post-Training v1 93M, UltraData-SFT 71M, FineVision 67M, Toucan-1.5M 36M, smoltalk2 36M, LLaVA-OneVision 23M, OpenCodeInstruct 17M, PixMo-ask 17M, AudioSkills-XL 15M, Video-R1 15M, AF-Chat / AF-Think 15M, LLaVA-Video QA 13M, VoiceAssistant-400K 9M (spoken question in), ToolACE 8M, xlam 7M, generation instructions 7M

sources.json has every source's tokens, documents, licence and removals; manifests/ names the exact original files and rows.

How pictures, clips and sounds become ids. A picture is 16 words, a clip 27 words (16 frames spread over the clip) with time stamps, a sound 50 words per 10-second window with time stamps. The words of every sample are already in the ids, so no media is needed to train. The codebooks that name new pictures, clips and sounds are in table/.

Sequence shapes

text                 <bos> text <eos>
picture + text       [image] 16 picture words [/image] caption
clip + text          [video] [t0] w w w w w [t2] w ... [/video] caption
sound + text         [audio] [t0] 10 words [t2] 10 words ... [/audio] labels / caption
generation           caption [gen_image] picture words [end_gen]     (and [gen_video], [gen_audio])
SFT                  <start_of_turn>user\n ... <end_of_turn>\n<start_of_turn>model\n ... <end_of_turn>
                     with [system] [think] [/think] [tools] [call] [/call] [result] [/result]

Decontamination. A document is removed when it contains at least half of the 13-grams of any test item of GSM8K, MATH, MMLU, ARC, HellaSwag, HumanEval or MBPP: 786 removed, all listed per source in sources.json. (Matching any single 13-gram, the common rule, removed 24-35% of maths and code for shared phrasing, not leaks.) The MSR-VTT 1k test clips are never used.

The table (table/)

94,514 rows x 512 bits: 65,536 language rows (the rest of the tokenizer is spelled with them; +0.9% length), 16,384 picture, 8,192 sound, 4,096 clip words, 4 markers, 302 control rows, and a unigram bias sealed over all 2.18B tokens. Picture, sound and clip words sit in language space: for CIFAR-10 test pictures the class's own word ranks at median 27 of 65,536 language rows (random would be ~32,768), for ESC-50 sounds at median 6. 500 language rows and 200 picture rows are held out of the data for the add-a-new-token test. The table is released as a fixed binary: usable, not rebuildable.

The experiment (training in progress)

body	ternary weights, int8 activations and an int4 KV cache inside training (straight-through), so the trained model and the CPU kernel compute the same thing
depth	2 prelude blocks, a tied core of 6 blocks run 2 times then 4 times from 60% of training, 2 coda blocks, boundary operator between passes
readout	correlation with the frozen table plus the sealed bias; text positions score language rows, positions inside a picture / sound / clip score that band
knowledge	Engram: hashed n-gram tables, ternary, trainable, outside the loop
long context	index branch trained from step one (KL to the model's own attention), sparse top-16 x 128-token blocks from 4k

Results, the checkpoint and the eval sheet are added here when the run finishes.

Scripts (scripts/)

pull.py (sources, with manifests), fetch_images.py, encode_words.py (pictures, clips, sounds to words), build_ids.py (sequences, decontamination), train_proof.py (the body).

Licences and credit

Each row carries its source's licence. Credit to the authors of every dataset above; CC-BY sources are attributed here and in sources.json. Some sources are non-commercial (AudioCaps, NVIDIA's AudioSkills-XL / AF-Chat / AF-Think under the NVIDIA OneWay Noncommercial License) and others state their own terms (MSR-VTT, Clotho, LLaVA-Video, FineVision's subsets): this release is for non-commercial research only, and use of each part is governed by its source's terms. A rights holder who wants material removed can open a discussion on this repository and it will be taken down.

Not included by design: how the frozen table is built, and the E8 eyes and ear.