This dataset is shared for research and for checking the Theia results. The photos belong to their owners (fashion-week coverage, brand websites, resale listings, auctions). By requesting access you agree to use them only for research and for checking these results, not to redistribute them, and to delete any photo on request. Requests are approved by hand.

Theia benchmark, version 1

A personal project, built on a small budget. This is the test side of the Theia launch at archiveagent.si, exactly as we ran it: the photos we tested on, the labels, every model's raw answers, the judge's verdicts, the protocol files and the code. Nothing here is re-scored or tidied up; the files are copies of the files we used.

Theia is Kimi K2.6 (Moonshot AI), fine-tuned with LoRA on Tinker (Thinking Machines). In the files it appears as ArchiveAgent v3 or kimi-k2.6-ft-A-m3. Not affiliated with or endorsed by Moonshot AI or Thinking Machines.

Both of our datasets are together in this Hugging Face collection.

All comparison models were run through OpenRouter (temperature 0, no tools, provider data collection denied). Theia was run on Tinker.

What is in the box

Folder	What
`images/`	the 1,699 test photos: `runway/` and `product/` (900 px WebP, named by the SHA-256 of the original file), `part4/` (the prepared JPEGs the models saw)
`tests/1_runway_test/`	287 photos, 41 shows, 14 houses: labels (`test_fresh.jsonl`), every model's answers (`answers/`, `answers_theia/`), run configs
`tests/2_runway_replicate/`	a second set of 572 photos, same kind of shows, no overlap with test 1; only Theia was run on it
`tests/3_blind_runway_and_product_house/`	Part 4: 50 runway photos and 150 product cases (75 products, one photo and a stack each): `labels.csv`, `cases.jsonl`, answers, `image_sources.jsonl`
`tests/4_exact_product/`	75 products, 428 photos: `test_pack.jsonl` (labels; `in_part4: true` marks the 75), answers, the judge's verdicts (`product-judgments.json`)
`analysis/`	the result files behind every number on the page: `all-tests.json`, `collab-fix.json`, `exact-all-systems.json`, and two checks
`code/`	the scoring and run code, as it is in our repository (it points at our private files; it is here to read)
`protocol/`	the plans and amendments, each with its SHA-256, written before the runs they cover
`prompts/`	the system prompts, schema and user text we sent
`training/`	what Theia was trained on, counts and hashes only: `training_image_sha256.txt` (7,512 hashes), `overlap_check.json`, the training log
`MANIFEST.sha256`	SHA-256 of every file in this folder

Results

Every missing, cut-off or "I don't know" answer counts as wrong. These tables are read from the files in analysis/.

Test 1, Runway Test: house, season and year all right (287 photos)

System	All three right	Share	Photos with no answer
Theia	162 of 287	56.4%	0
Kimi K3	129 of 287	44.9%	93
Claude Opus 5.5	82 of 287	28.6%	71
GPT-6.1 Sol	75 of 287	26.1%	26
Gemini 3.1 Pro	66 of 287	23.0%	105
Claude Sonnet 5.5	62 of 287	21.6%	116
Kimi K2.6	35 of 287	12.2%	128
Qwen3.5-397B	15 of 287	5.2%	82
Grok 4.7	4 of 287	1.4%	173
Inkling-Small	3 of 287	1.0%	255
Qwen3.6-35B-A3B	1 of 287	0.3%	98
Qwen3.5-9B	0 of 287	0.0%	242

Test 3a, Product House Test: the house named (150 cases)

System	Right	Share	No answer
Theia	150 of 150	100.0%	0
Claude Opus 5.5	148 of 150	98.7%	2
GPT-6.1 Sol	143 of 150	95.3%	6
Claude Sonnet 5.5	139 of 150	92.7%	8
Gemini 3.1 Pro	83 of 150	55.3%	67
Grok 4.7	31 of 150	20.7%	116

Test 3b, Blind Runway Check: all three right (50 photos)

System	Right	Share	No answer
Theia	28 of 50	56.0%	0
Gemini 3.1 Pro	13 of 50	26.0%	14
Claude Opus 5.5	11 of 50	22.0%	14
Claude Sonnet 5.5	11 of 50	22.0%	24
GPT-6.1 Sol	10 of 50	20.0%	3
Grok 4.7	0 of 50	0.0%	37

Test 4, Exact Product Test: the exact product named (75 products; a second model judges)

System	Right	Share	Note
Claude Opus 5.5	44 of 75	58.7%	run on 75 of 75
GPT-6.1 Sol	39 of 75	52.0%	run on 75 of 75
Theia	38 of 75	50.7%	run on 75 of 75
Grok 4.7	34 of 75	45.3%	run on 75 of 75
Claude Sonnet 5.5	27 of 75	36.0%	run on 54 of 75, the rest not run: lower bound
Gemini 3.1 Pro	-	-	run on 39 of 75, every answer cut off, so no score

Check there is no overlap with training

training/training_image_sha256.txt has the SHA-256 of every one of the 7,512 photos Theia was trained on. Every test photo has a SHA-256 of its original file in the test labels (sha, shas) or in image_sources.jsonl. training/overlap_check.json shows the two lists share nothing (1,477 test hashes checked). If you hold the original photos, you can hash them and check yourself. We also ran a near-duplicate check with SigLIP embeddings privately; that needs the training photos, which are not released.

More benchmarks, run on a local computer

Before Theia we ran a separate set of benchmarks on four open image models, on one home computer with an NVIDIA RTX 5060 (8 GB of VRAM): how well each finds the right fashion house in 86,874 runway photos, how long each took to embed them, and six more tests. That repository holds protocols, results, reports and code, and no photos. If you want to understand those benchmarks, request access here. It works the same way as this repository: the card is visible and I approve requests by hand.

Limits, said plainly

One run per system. Labels were made by models and trusted by the owner; the exact-product judge is a model.
Theia always answers, even when unsure. It was trained on 14 houses.
Claude Opus 5.5 is ahead of Theia on the Exact Product Test.
Claude Sonnet 5.5 and Gemini 3.1 Pro were not finished on the Exact Product Test because the testing budget ran out. Sonnet's number is a lower bound. Gemini was run on 39 of 75 and every answer was cut off at the token limit, so it has no score (a dash).
Product photos in the test files carry marketplace group ids (gr-...) exactly as in our files.
This is version 1 and it is shared with anyone I approve. We cannot control what they train on, so it can no longer be a clean test. A fresh hidden set is the plan for the next version.

Access

The photos and files are behind a request button; the card and the result tables above are visible to everyone. Requests are approved by hand. You can also email kalanpeace@gmail.com and say what you are working on.

What is not here

The training photos, the training rows, the 86,874-photo archive, embeddings and the Theia weights. If you are a lab or want the full archive, more training detail or a walkthrough, email kalanpeace@gmail.com.

Licence

Our labels, answers, scores and analysis: CC BY 4.0. Our code: Apache-2.0. The photos are not ours. They come from fashion-week coverage, brand websites, resale listings and auctions and belong to their owners; they are shared here for research and for checking our results. If you own a photo and want it removed, email kalanpeace@gmail.com and it will be taken down.

Downloads last month: 321

Total file size:

112 MB

Collection including kalanpeace/theia-bench

Theia and ArchiveAgent benchmarks

Collection

Theia, a fashion model, and its benchmarks. Request-access datasets: the Theia test set, and local image-model tests. A personal project. • 2 items • Updated Oct 9