Card: fact-check corrections (sources, PixArt compute, caption fields, dataset filtering wording, param count with TAESD vs SD-VAE)

README.md

@@ -77,7 +77,7 @@ shows all of that in pictures, along with every model we trained on the way to A
   generator, 512-token context.
 - **Decoder:** [SD-VAE-ft-MSE](https://huggingface.co/stabilityai/sd-vae-ft-mse), or
   [TAESD](https://huggingface.co/madebyollin/taesd) for speed.
-- **Data:** [FLUX-Reason-6M](https://huggingface.co/datasets/LucasFang/FLUX-Reason-6M), 5,654,461 of its images, each with 8
+- **Data:** [FLUX-Reason-6M](https://huggingface.co/datasets/LucasFang/FLUX-Reason-6M), 5,654,461 of its 5,890,279 images, each with up to 8
   English captions each. **Compute:** Arrhenius (EuroHPC, NAISS, Sweden), NVIDIA GH200.
 - **Licence:** MIT, for the code and the weights.
 
@@ -141,11 +141,18 @@ what each model cost to train:
 | SDXL | 2.6B UNet | not published | 0.55 |
 | Meissonic | 1B | ~1,150 H100-h | 0.54 |
 | SD 2.1 | 865M UNet | 200,000 A100-h | 0.50 |
-| PixArt-α | 0.6B (+4.3B T5) | ~16,000–18,000 A100-h | 0.48 |
+| PixArt-α | 0.6B (+4.3B Flan-T5-XXL) | ~753 A100-days (~18,000 A100-h) | 0.48 |
 | SD 1.5 | 860M UNet | 150,000 A100-h (SD 1.x family) | 0.43 |
 
 The A100 equivalence uses a GH200 : A100 wall-clock ratio of 2.5–3.2×.
 
+Sources:
+- GenEval for SD 1.5, SD 2.1 and SDXL: the GenEval paper (arXiv 2310.11513, Table 2).
+- PixArt-α: the SD3 paper (arXiv 2403.03206, Table 5).
+- Meissonic: arXiv 2410.08261, Table 3.
+- Compute figures: each model's card or paper. PixArt-α's 753 A100-days are converted from 1,656 V100-days; an
+  earlier version of that paper said 675.
+
 <p align="center"><img src="assets/geneval_categories.png" alt="GenEval by category for Agate, SD 1.5 and Supra2-IMG" width="100%"></p>
 
 **Every model we trained, same protocol (1 image per prompt, 553 images each):**
@@ -170,7 +177,7 @@ way, Agate scores 0.535 and 0.550.
 
 ### Qwen-Image-Bench
 
-[Qwen-Image-Bench](https://huggingface.co/Qwen/Qwen-Image-Bench) asks its judge model, Q-Judger (27B), to grade
+[Qwen-Image-Bench](https://huggingface.co/Qwen/Qwen-Image-Bench) asks its judge model, Q-Judger (fine-tuned from Qwen3.6-27B), to grade
 each image against a per-prompt checklist in five areas. We ran the judge exactly as published: their checklists
 and aggregation code, greedy decoding, thinking on.
 
@@ -293,15 +300,18 @@ noise latent 4×32×32 ─┬─► thinker: prelude(3) → core(3) × 4 loops
 | Caption phase, steps 44k – 99.25k (to epoch ~26) | 2 nodes × 4 GH200 (8 GPUs), 7 h 32 min, batch 1,024, 512 text tokens, caption mix, constant lr then cosine to 2e-5 (10 % floor) over the last 16.5k steps |
 | Total | 99,250 steps, 146.9M images seen, 144.7 GH200-hours |
 
-**Caption mix.** FLUX-Reason-6M ships eight English captions per image. In 80 % of batches each image draws one
-short caption, weighted toward *composition* (layout is the weakest GenEval skill). The other 20 % use every
+**Caption mix.** FLUX-Reason-6M has eight English caption fields: composition, entity, text, imaginative, style,
+abstract, original and detail. Not every image has every field; each image draws from the fields it has.
+In 80 % of batches each image draws one short caption, weighted toward *composition* (layout is the weakest GenEval skill). The other 20 % use every
 image's long *detail* caption (median 217 tokens). The model therefore learns both what people type and what
 carries the concepts, and the jump to 512 tokens stops long captions from being truncated.
 
 ## Training data and safety
 
-**Why FLUX-Reason-6M.** We chose the training set because it is curated and contains no adult images. Agate never
-saw explicit material during training, so it has no learned way to produce it.
+**Why FLUX-Reason-6M.** We chose a curated, fully synthetic dataset built for compositional and reasoning prompts:
+objects, layouts, text, styles and imaginative scenes, not people or adult content. Its authors document
+quality filtering (blur, noise, structural distortion) but do not describe an NSFW filter. We have not audited
+all 5.65M images, so we cannot promise it contains no adult images. We expect such content to be rare at most.
 
 **Red-teaming.** Before release we tested Agate adversarially for sexual content and for child sexual abuse material
 (CSAM).
@@ -309,8 +319,8 @@ saw explicit material during training, so it has no learned way to produce it.
 - **Adult nudity:** we could not get Agate to generate genitals in any attempt. The worst output was a blurry nude
   female figure with no anatomical detail (no areolae).
 
-We consider explicit or abusive output practically infeasible for this model, because it was never trained on
-such material, and misuse unlikely. That is our assessment, not a guarantee. Agate ships without a safety
+Given the training data and these results, we consider explicit or abusive output practically infeasible
+for this model, and misuse unlikely. That is our assessment, not a guarantee. Agate ships without a safety
 classifier, so public-facing deployments should add one.
 
 Using Agate, or any model derived from it, to depict minors sexually, to create non-consensual intimate imagery,