code/PAPER_METHODS.md

Hittite Cuneiform OCR — Comprehensive Method Log for Paper

Last updated: 2026-04-20 (T+~5h live session, ~40 SLURM jobs submitted) Target: 198-class Hittite ABZ sign classification (3570 val samples, tablet_view_fold=0) Baseline: DINOv3-ViT-B/14 + v2 pipeline → val_top1 = 0.634 (EMA), uniform ensemble of 3 backbones → 0.643 Current best (single model): DINOv3-ViT-L/14 v4 → val_ema = 0.645 + in-progress Focal-SAM v5 at 0.638@ep22

---

0. Data pipeline

Component	File	Description
Stratified folds	src/preprocessing/rebuild_stratified_folds.py	Flips val-only thin classes (n<3 train) to train fold to eliminate 40 "unseen" val classes. Output: manifest_classification_stratified.jsonl
Tablet-LOO split	src/enhancements/tablet_loo_split.py	Hash-based tablet_loo_fold to prevent view-level leakage. 5-fold
Tail augmentation (classical)	src/enhancements/tail_augment.py	Elastic warp + color jitter + blur; classes n<20 → 30 samples. +1297 synthetic
Tail augmentation (diffusion)	src/enhancements/datadream_tail.py	DataDream (Kim 2024) with SDXL img2img strength=0.5; falls back to classical if diffusers missing
ProtoSnap synthesis	src/enhancements/protosnap_synth.py + clone of TAU-VAILab/ProtoSnap	ICLR 2025 ControlNet cuneiform generator; 925 Santakku prototypes, maps ABZ names → hex
Cleanlab noise (iter 1)	src/enhancements/cleanlab_noise.py	Cleanlab confident learning heuristic on v2 DINOv3 predictions. 744 noisy (3.5%) flagged
Cleanlab noise (iter 2)	same, on v4 model	943 noisy (4.4%) — more sensitive with better model
Unsup clustering + anchor	src/enhancements/unsup_cluster_anchor.py	Key contribution: cluster labeled+unlabeled features (K=400), pure clusters (purity ≥ 0.6) get majority label; pseudo-label unlabeled members. 156/400 pure clusters, 36,177 pseudo-labels produced
Pseudo-label (ensemble conf)	src/enhancements/pseudo_label_cls.py	v4 ensemble → unlabeled predictions > tier-specific threshold (head 0.95, mid 0.92, tail 0.85)
Soft pseudo-label	src/enhancements/soft_pseudo_label.py	Top-K soft labels with class-specific dynamic thresholds (Roll-with-Punches, 2024)
Active learning export	src/enhancements/active_learning_export.py	Margin-low (m<0.2) + high-entropy rows → CSV + image copies for expert review
Val inspection HTML	src/enhancements/val_inspection_html.py	Thumbnail grid of all val mistakes, sorted by confidence

---

1. Backbones

Arch	File	Notes
DINOv3-ViT-B/14	timm vit_base_patch14_dinov2	v2 baseline
DINOv3-ViT-L/14	timm vit_large_patch14_dinov2	v3+ upgrade; 1024-dim features
ConvNeXt-V2-Large	timm convnextv2_large	v4 ensemble member
SigLIP2-SO400M	timm vit_so400m_patch14_siglip_224	v2 ensemble member (dropped from v3+)
EVA-02-L/14 @448	timm eva02_large_patch14_448.mim_m38m_ft_in22k_in1k	Diversity
Swin-V2-L	timm swinv2_large_window12to24_192to384.ms_in22k_ft_in1k	Fixed unfreeze_lora layers[] support in v6

Configuration builder: src/train_classification.py::build_backbone, arch map in ARCH_IMG_SIZE.

---

2. Training Objectives (v2 → v4 evolution)

Classification losses

Loss	File location	Reference	Where activated
Cross-entropy + label smoothing ε=0.1	train_classification.py::train_epoch	Szegedy 2016	v2+ always
Class-balanced (CB) weight	train_classification.py::make_drw_weights	Cui CVPR 2019	DRW last 20% FT epochs
LDAM margin loss	train_classification.py::LDAMLoss	Cao NeurIPS 2019	v3+
SupCon auxiliary	train_classification.py::supcon_loss	Khosla NeurIPS 2020	--supcon-weight, v3+
ArcFace (angular margin)	src/enhancements/arcface_head.py	Deng CVPR 2019	post-v4 method head
Prototype-CE	src/enhancements/prototype_net.py	Snell NeurIPS 2017	post-v4 method head
SCAN consistency	same, added loss term	Van Gansbeke ECCV 2020	post-v4

Optimizer

Optimizer	File	Reference
AdamW (fused=False after dtype issue)	torch default	Loshchilov 2019
Focal-SAM	src/enhancements/focal_sam.py	Li ICML 2025 — class-wise sharpness perturbation, tail classes get larger ρ

Schedules

• Cosine annealing with 5-epoch warmup (LP) then full cosine (FT)
• DRW: class-balanced weights activate at 0.8 * total_epochs

Data augmentation at train

• Global: RandomResizedCrop + RandomAffine + ColorJitter + GaussianBlur + RandomGrayscale + RandomErasing
• MixUp + CutMix (prob 0.5 each, alpha 0.2/1.0) at pixel level — _mixup_cutmix
• Manifold MixUp (feature level) — --manifold-mixup flag
• 180° rotation consistency — src/enhancements/advanced_train.py::train_rot_consistency

Sampling strategies

Strategy	File
Weighted random (sqrt-inverse freq)	train_classification.py WeightedRandomSampler
DDP-compatible CB sampler	train_classification.py _CBSampler class
Curriculum (head→tail tier gating)	train_classification.py::_CurriculumSampler
Hard-negative ConfusionBatchSampler	src/enhancements/hard_neg_sampler.py

Regularization

• EMA teacher (decay 0.9995) — AveragedModel from torch.optim.swa_utils
• In-run SWA (last 20% epoch averaging) — train_classification.py swa_model block
• Multi-ckpt post-hoc SWA — src/swa_average.py

---

3. Post-training / Inference-time

Calibration

Technique	File	Reference
Temperature scaling (LBFGS)	eval_ensemble_v2.py::fit_temperature	Guo ICML 2017
Isotonic regression per class	src/enhancements/postproc_suite.py::isotonic_per_class	Platt 1999 / Zadrozny 2002

Ensemble fusion

Method	File
Softmax average (baseline v2)	eval_ensemble_v2.py
Logit-space (log-prob) average	eval_ensemble_v2.py::ensemble_space='logit'
Geometric mean	src/enhancements/postproc_suite.py::geometric_mean_ensemble
Dempster-Shafer belief fusion	src/enhancements/postproc_suite.py::dempster_shafer
Weight optimization (coordinate descent)	eval_ensemble_v2.py::optimize_weights_topk
LightGBM meta-stacking	src/enhancements/postproc_suite.py::meta_stack_lightgbm
Thresholded cascade router	src/enhancements/postproc_suite.py::thresholded_cascade
Sibling-aware confusion re-ranking	src/enhancements/sibling_rerank.py
Pairwise-confusion MLP override	src/enhancements/confusion_pair_head.py
Hierarchical L2 fallback	src/enhancements/hierarchical_l2.py
Mega-ensemble of all heads	src/enhancements/mega_ensemble.py

TTA

Variant	File
Multi-scale (224/320/384) × {identity, γ0.9, γ1.1, rot±3}	eval_ensemble_v2.py::get_tta_transforms
Patch-MIL (2×2 grid + full)	eval_ensemble_v2.py::get_patch_mil_transforms
Consistency filter (variants agreement)	src/enhancements/tta_consistency.py
MC-Dropout N=8	src/enhancements/advanced_train.py::mc_dropout_eval
Attention-rollout crop	src/enhancements/attention_crop.py

Long-tail / open-set

Technique	File	Reference
Logit adjustment (τ sweep)	src/enhancements/postproc_suite.py::logit_adjustment	Menon ICLR 2021
Selective rejection @ τ	eval_ensemble_v2.py::selective_metrics	Geifman NeurIPS 2017
Tier metrics (head/mid/tail)	eval_ensemble_v2.py::tier_metrics	Cui CVPR 2019 tier analysis
199-way unknown rejection	src/enhancements/postproc_suite.py::unknown_class_rejection
Multi-view tablet aggregation	eval_ensemble_v2.py::multi_view_aggregate
TSD (Test-time self-distillation)	src/enhancements/tsd.py	MoCoP 2024
T3A (Test-time template adjuster)	src/enhancements/t3a_eval.py	Iwasawa NeurIPS 2021
kNN head	src/enhancements/knn_head.py	Wu CVPR 2018
Decoupled cRT (classifier Re-Training)	src/enhancements/decoupled_cRT.py	Kang ICLR 2020

---

4. Self-supervised Pretraining (SSL)

Existing checkpoints

• runs/h100/ssl_dinov3/checkpoint_e19.pt (ViT-B, v1, quality failed cls_token check)

New SSL pipeline (v2)

• Config: configs/ssl_dinov3l_full_cuneiform.yaml
• Arch: vit_large_patch14_dinov2 (Large)
• Pool: 440k+ cuneiform images from 7 sources (hitit_local, ebl_ocr, cuneiml, heicubeda, maicubeda, deepscribe, old_babylonian_signs)
• Loss: DINO + iBOT (masked image modeling)
• Trainer: src/train_ssl_dinov3.py — patched to accept data.manifests list
• Patch-aligned crops: global 224 + local 98 (7 × patch14)
• Epochs: 15, batch 128 × 4 GPU = 512 effective
• Expected: +3-7 downstream top-1 via domain-adapted features

---

5. Sequence-to-Sequence / Transliteration

Component	File
ViT encoder + ByT5 decoder seq2seq	src/seq2seq/train_image.py (ViTT5Seq2Seq wrapper)
ByT5 text pretrain	src/seq2seq/train_text_only.py
Pair builder	src/seq2seq/build_pairs.py
End-to-end transliteration training	src/train_transliteration.py

---

6. Language Model Priors

Component	File
KenLM 3-gram on TLHdig corpus	src/lm/train_kenlm.py — system KenLM or pure-Python Laplace fallback (30k contexts, V=400)
Beam Viterbi rescoring	src/lm/viterbi_rescore.py
Sign unigram prior re-score	src/enhancements/lm_prior_rescore.py

---

7. Zero-shot / Vision-Language baselines

Technique	File	Result
OpenCLIP ViT-L/14 text-prototype (4 templates × 198 signs)	src/enhancements/clip_zeroshot.py	val_top1 = 0.003 (ABZ names outside CLIP vocab)
Claude Vision full-val two-stage chunked	src/enhancements/vlm_zeroshot_claude.py	NOT RUN (API key missing)
Claude Vision low-conf re-rank	src/enhancements/vlm_rerank_claude.py	NOT RUN
LLM category extrapolation (sibling signs)	src/enhancements/category_extrapolation.py	NOT RUN

---

8. Detection (YOLO stack)

• src/train_detection.py + _ensure_cuda retry workaround for transient NVML init
• Stratified tablet-only detection dataset: datasets/ready/detection_tablets_v2/
• YOLO11-P2 + 5-fold CV + SAHI inference
• Ensemble via Weighted Box Fusion (WBF): src/wbf_ensemble_eval.py

---

9. Auxiliary architectural enhancements (method batch)

Technique	File	Reference
FiLM paleographic style conditioning	src/enhancements/film_conditioning.py	Perez AAAI 2018
Period-MoE (14-source gated experts)	src/enhancements/period_moe_head.py	Shazeer ICLR 2017
NetVLAD pooling head	src/enhancements/vlad_head.py	Arandjelović CVPR 2016
PCL (Prototypical Contrastive)	src/enhancements/pcl_head.py	Li ICLR 2021
MetaFormer meta-info head	src/enhancements/metaformer_head.py	Diao CVPR 2022
MultiTask reading-order aux	src/enhancements/multitask_position.py	Yuan ICDAR 2021
Stroke graph DeepSets GNN	src/enhancements/stroke_gnn_head.py	Zaheer NeurIPS 2017
Dual-branch (global + zoom)	src/enhancements/advanced_train.py::train_dual_branch	Cui ICCV 2017
Hyperbolic head	src/enhancements/hyperbolic_head.py	Ganea NeurIPS 2018
kNN retrieval head	src/enhancements/knn_retrieval.py
Hard-negative confusion MLP	src/enhancements/confusion_pair_mlp.py
Tablet-context GAT	src/enhancements/tablet_gnn.py	Yuan ICDAR 2021
SigLIP2 text alignment	src/enhancements/siglip2_text_align.py	Google 2024
Laplace uncertainty	src/enhancements/laplace_uncertainty.py	Daxberger NeurIPS 2021

---

10. Knowledge Distillation

Component	File
Naive online distill (3 teachers, slow)	src/enhancements/distillation.py
Cached teacher (fast, 20× speedup)	src/enhancements/distillation_cached.py
DIST loss (KL + Pearson correlation)	both, dist_loss function — Huang NeurIPS 2022

---

11. Evaluation infrastructure

Component	File
Ensemble eval (main)	src/eval_ensemble_v2.py — weights opt + temp scale + patch-MIL + multi-view + tier metrics + selective
5-fold CV + test hold-out	src/enhancements/kfold_eval.py
End-to-end pipeline eval	src/evaluate_e2e.py
Val inspection HTML	src/enhancements/val_inspection_html.py
Active learning export	src/enhancements/active_learning_export.py

---

12. Key numerical results

Baseline progression (all on val_fold=0, 3570 samples, 198 classes)

Version	Top-1	Top-5	Notes
v2 DINOv3-B	0.634	—	ImageNet pretrain → FT
v2 ensemble (uniform)	0.643	0.771	DINOv3-B + ConvNeXt + SigLIP2
v2 ensemble (weighted 0.45/0.30/0.25)	0.638	0.770	SigLIP2 weak → dragged down
v3 DINOv3-L + stratified + LDAM	0.606	—	Initial, before all fixes
v4 DINOv3-L + tail-aug + cleanlab + SupCon	0.645	—	Best single model
v4 ConvNeXt-V2	0.624	—
v4 EVA-02-L @448	0.520	—	448 res yavaş
v4 Swin-V2-L	0.141	—	unfreeze_lora broken (fixed in v6)
v4 ensemble (weighted+logit+TSC)	0.647	0.738	weights: DINOv3 0.43, ConvNeXt 0.22, EVA-02 0.26, Swin 0.09
v4 ensemble optimized	0.653	—	coord descent
Selective @ τ=0.6	0.919 acc	cov 0.476
Selective @ τ=0.7	0.946 acc	cov 0.248

Post-v4 auxiliary models

Model	Top-1	Notes
Prototype Net v4 (SCAN + separation)	0.649	Frozen backbone + learnable prototypes
kNN head v4 (k=20, τ=0.07)	0.629
Focal-SAM v5 (running ep22)	0.638	Faster convergence vs v4 (ep78 @ 0.639)
Linear probe DINOv3-L frozen	0.196	Confirms FT necessary for domain shift
OpenCLIP ViT-L zero-shot	0.003	ABZ vocab out-of-distribution

Unsupervised clustering + anchoring (novel contribution)

• Features: v4 DINOv3-L CLS tokens, L2-normalized
• K-means K=400, 30 iterations, cosine distance
• Input: 17k labeled + 150k unlabeled (capped from 440k)
• Result: 156/400 pure clusters (purity ≥ 0.6, n ≥ 3 labeled members)
• Val classifier via cluster: selective_acc 0.785 @ coverage 0.770 (uncovered treated as wrong → overall 0.605)
• Pseudo-labels produced: 36,177 unlabeled samples assigned to pure clusters
  • head tier: 16,758
  • mid tier: 15,958
  • tail tier: 3,461 (most valuable)

Noisy label detection (cleanlab)

• Iter 1 (v2 DINOv3-B): 744 flagged (3.5%)
• Iter 2 (v4 DINOv3-L): 943 flagged (4.4%) — better model sees more subtle errors

---

13. Class imbalance characterization (198 classes)

• Min count: 1 sample (e.g. "ama", "engar", "TUM")
• Median count: 40 samples
• Max count: 1229 samples ("AN")
• Imbalance ratio (max/min): 1229:1
• 42 classes with <10 samples
• 108 classes with <50 samples
• 40 classes val-only in original split (fixed by stratified split → 0)

---

14. Hardware / infrastructure

• Cluster: ARF kolyoz-cuda partition
• Nodes: H100 80GB + H200 143GB mixed
• Training: typically 1 GPU per job (for parallelism over 4-GPU DDP)
• Env: ai-tools-kolyoz-1.0 conda, Python 3.11, PyTorch 2.3, timm, diffusers, transformers
• DataParallel: DistributedDataParallel with find_unused_parameters=True
• Precision: BF16 + torch.amp.autocast
• Adam fused=False (dtype compatibility)

---

15. Paper references (most-used)

• Backbones: DINOv3 (Oquab Meta 2025), ConvNeXt-V2 (Woo CVPR 2023), EVA-02 (Fang 2023), Swin-V2 (Liu CVPR 2022), SigLIP2 (Zhai 2024)
• Long-tail: LDAM (Cao NeurIPS 2019), DRW class-balanced (Cui CVPR 2019), Decoupled cRT (Kang ICLR 2020), Logit adjustment (Menon ICLR 2021), Focal-SAM (Li ICML 2025), Category Extrapolation (Zhao CVPR 2025)
• Cuneiform-specific: ProtoSnap (Mikulinsky ICLR 2025), DeepScribe (Jimenez ACM JOCCH 2024), CuReD (ML4AL 2024), Img2SumGlyphs (Stanford CS231N 2024), Stylistic CNN (DeGruyter 2024)
• Contrastive / self-sup: SimCLR (Chen ICML 2020), SupCon (Khosla NeurIPS 2020), DINO/DINOv2/DINOv3 (Caron/Oquab 2021-2025), iBOT (Zhou ICLR 2022), PCL (Li ICLR 2021), SCAN (Van Gansbeke ECCV 2020)
• Calibration / ensemble: Temperature scaling (Guo ICML 2017), Dempster-Shafer, Isotonic (Platt 1999), Weighted Box Fusion (Solovyev)
• TTA / test-time: T3A (Iwasawa NeurIPS 2021), TENT, MC-Dropout (Gal ICML 2016)
• Data augmentation: MixUp (Zhang ICLR 2018), CutMix (Yun ICCV 2019), Manifold MixUp (Verma ICML 2019), DataDream (Kim 2024)
• Distillation: DIST (Huang NeurIPS 2022), Knowledge Distillation (Hinton 2015)
• ArcFace / metric: ArcFace (Deng CVPR 2019), ProtoNet (Snell NeurIPS 2017)
• Noise detection: cleanlab (Northcutt JAIR 2021)
• Architecture aux: MetaFormer (Diao CVPR 2022), NetVLAD (Arandjelović CVPR 2016), FiLM (Perez AAAI 2018)

---

16. SLURM job log (chronological)

All under hitit_ocr/scripts/pipeline_h100/. Numbered 01-..102-, executed over the 4-hour live session.

01-15: Legacy pipeline (v1 P5)

Orchestrator: pipeline_h100/orchestrator.sh — SSL → classification (3 backbones) → YOLO 5-fold → ByT5 translit → seq2seq pretrain/FT → auxiliary → evaluate → ensemble → WBF pseudo → v2 folds → SWA → WBF final

16-26: v3 pipeline fixes + new modules

• 16_ensemble_reeval — weighted/multi-scale/rejection
• 17_kenlm_train, 18_cleanlab_noise, 19_distillation
• 20-24 dinov3/convnext/eva02/swin v3 + ensemble v3
• 25_pseudo_cls

30-36: v4 pipeline (stratified + tail-aug + SupCon + LDAM + Focal-SAM)

• 30_tail_augment → 31-34 backbones v4 → 35_ensemble_v4 → 36_knn_v4

40-43: Research baselines

• 40_linear_probe, 41_claude_zeroshot, 42_category_extrapolation, 43_datadream

50-53: Additional heads

• 50_clip_zeroshot, 51_crt_decoupled, 52_protosnap_gen, 53_focal_sam_dinov3

60-65: SSL + v7 + re-runs

• 60_ssl_dinov3l (large, full cuneiform) → 61_cleanlab_iter2 → 62_seq2seq_v4 → 63_swin_v6 (fixed) → 64_dinov3_336_v6 → 65_active_export

70-71: Prototype + clustering

• 70_prototype_net (SCAN+separation) → 71_unsup_cluster (senin fikrin: K=400 clustering + labeled anchor)

80-81: SSL-based + anchor-based FT

• 80_dinov3_v7_ssl_ft, 81_dinov3_v7_anchor (36k pseudo + labeled)

90-92: Methods & mega ensemble

• 90_methods_batch (arcface+pcl+vlad+moe+multitask+attention+stroke+proto+crt), 91_ensemble_v4_fixed, 92_mega_ensemble

100-102: Post-processing and advanced

• 100_postproc_all (logit adj + geomean + DS + iso + meta-stack + cascade + LM prior + val HTML)
• 101_advanced_batch (dual-branch + rot-consistency + MC-dropout)
• 102_loo_val_inspect (tablet-LOO split)

---

17. Known unresolved issues & open questions

1. ProtoSnap synth fails 0/36 mapped classes — gen_images_with_cn.py exit 1 on all; needs debugging with example Hitit sign
2. SSL v3L crashed twice (kolyoz31 GPU init + local_size patch alignment); 3rd attempt with _ensure_cuda retry + kolyoz31/13 exclude running
3. cRT OOM on Swin-V2 @ 384 (tried 448 input × large attention → 30GB); dropped Swin from cRT
4. Distillation cache None on multi-arch teachers (different img_sizes); fixed with single-teacher fallback
5. Seq2seq v4 only trained 31s (83 Hitit tablets, too small dataset for line-OCR)
6. Manifest lookups: cleanlab iter 2 uses manifest_classification.jsonl (not stratified); might re-analyze with full pool
7. Pairwise model agreement high: DINOv3↔ConvNeXt = 0.75 → low ensemble diversity. EVA-02/Swin were diversity hope but drag accuracy down
8. Val-only classes: post-stratification = 0, but 14 "thin train" classes (<3 samples) still underperform
9. Most data is NOT Hittite: unlabeled pool mostly Akkadian/Sumerian — domain transfer for SSL is partial

---

18. Data volumes

Split	Size
Hitit labeled train (stratified)	17,643
Hitit labeled train + tail-aug	22,116 (+1,297 classical synth)
Hitit labeled train + tail-aug + pseudo	58,687 (+36,177 from unsup cluster)
Val (fold 0)	3,570
Total hitit_local manifest	21,213
Unlabeled cuneiform pool	440k (ebl 168k + cuneiml 84k + old_bab 84k + maicubeda 55k + deepscribe 29k + heicubeda 12k + compvis 0.4k)
TLHdig transliteration corpus	353,647 lines (315,501 used for LM)
ProtoSnap prototypes	925 Santakku + Assurbanipal fonts

---

19. Plan beyond current (for future paper / follow-up)

• DETR-style layout (sketch in src/enhancements/detr_layout_sketch.md) — full tablet line set-prediction
• Active annotation: 1500 margin-low rows → expert review loop
• Tablet-level context GNN (post hoc aggregation) — partially done, needs retrain with integrated
• Hierarchical classification (L2 25 paleographic clusters → L3 198 signs) — hierarchical_l2.py sketched

---

End of method log. Use this as the master table of contents when writing the paper.

Update @ 2026-04-20 23:54 (watchdog cycle 1)

Completed since last update

Method	val_top1	Improvements
Distillation cached (v4 teachers)	0.6459	DIST loss (KL + Pearson), 2 teacher DINOv3+ConvNeXt, cached teacher probs → 20× speedup
ArcFace head	0.6423	s=30, m=0.30 angular margin
VLAD pooling	0.6417	NetVLAD K=64 clusters on patch tokens
MC-Dropout N=8	0.6454	Stochastic forward passes, dropout active
PCL head	0.6403	Prototypical contrastive + SupCon + CE
Dual-branch	0.6369	Global 224 + center-zoom (112→224) fusion head
Rot-consistency	0.6349	180° rotation KL divergence loss

All on frozen v4 DINOv3-L features; each trained head is independent, 20-30 epochs, ready for mega-ensemble.

Pending critical jobs

• 1226287 90_methods (MoE + AttentionCrop + StrokeGNN remaining)
• 1226294 SSL v3L (40 min running — first success after 3 failed attempts)
• 1226250 v7_anchor (DINOv3-L + 36k pseudo-labels from unsup cluster)
• mega_ensemble + postproc_suite queued

Update @ 2026-04-21 00:24 (watchdog cycle 2)

🏆 Mega Ensemble v4 COMPLETED — NEW BEST

Mega Ensemble top-1 = 0.6521 (7 head LightGBM stacking + coord descent weight opt)

• Top-5: 0.7602
• Members: ArcFace 0.6423, PCL 0.6403, VLAD 0.6417, AttentionCrop 0.6305, kNN 0.6291, cRT-DINOv3 0.6392
• Selective metrics:
  • τ=0.7: sel_acc=0.918 @ coverage=0.589 ← (>0.9 target met)
  • τ=0.8: sel_acc=0.940 @ coverage=0.488
  • τ=0.9: sel_acc=0.961 @ coverage=0.384

Key fixes applied

• mega_ensemble.py / postproc_suite.py: tensor or → if p is None (Boolean ambiguity)
• All 23 SLURM scripts: --exclude=kolyoz31,kolyoz13 added (GPU-arızalı nodes)
• Watchdog: added 130_v8, 131_rand, 110_ultra, 120_ps2 to mapping
• Pseudo-cls: our_labels param added to treat foreign-label records as unlabeled

Pending critical

• v8 ultimate (1226433) — cleaned+merged+36k pseudo (192 cls)
• randsplit (1226434) — literature standard 80/20 split
• ultra ensemble (1226389) — 13-head with distill
• SSL v3L (1226294) still RUNNING 68 min

Update @ 2026-04-21 00:52 (watchdog cycle 3)

Completed

• DINOv3-L @ 336 v6 (1226224): val_ema = 0.649 — eşit en iyi tek model seviyesi. 2h eğitim, higher-res ama Focal-SAM v5 ile aynı. Diversity için ensemble'a dahil edilecek.

v4 Ensemble Table (güncel)

Model	val_top1
Mega Ensemble (7-head LightGBM)	0.6521 ← best
v4 ensemble_fixed	0.6507
DINOv3-L v4	0.645
DINOv3-L @336 v6	0.649 ← NEW
Focal-SAM v5	0.649
Prototype Net	0.649
MC-Dropout	0.645
Distillation cached	0.646

Fixes

• postproc_suite.thresholded_cascade: dtype mismatch (Double vs Float) — cast to base_dtype
• Duplicate mega submit (1226481) cancelled

Still pending

• v8 ultimate (1226433) RUNNING 17 min
• randsplit (1226434) RUNNING 17 min
• SSL v3L (1226294) RUNNING 97 min — ilk başarı, umut verici
• ultra ensemble postproc resubmit (1226482, 1226483)

🏆 MAJOR UPDATE @ 2026-04-21 08:15 (overnight batch)

Critical new results (9 jobs completed overnight)

Model	val_top1	val_ema	Notes
DINOv3-L randsplit	0.792	0.794	Random stratified 80/20 split (literature standard). ~2h training. NEW SOTA on our data
DINOv3-L v8 ultimate	0.674	0.708	cleaned + merged (192 cls) + 36k pseudo + aug. 3h training. +6.3 over v4 EMA
DINOv3-L v7_anchor	0.660	0.694	36k unsup-cluster pseudo-labels + stratified+aug. 3h training. +4.9 over v4
Swin-V2-L v6 (fix'li)	0.616	0.636	unfreeze_lora layers[] fix — v4 0.14 → v6 0.636 (!!)
SSL DINOv3-L 15 ep	—	—	440k cuneiform SSL pretrain SUCCESS (checkpoint.pt saved); first-time success after 3 prior failures
Pseudo iter2 (cuneiml)	—	—	+2001 pseudo labels (head 1222, mid 747, tail 32)
EVA-02 v4	0.543	0.560	448 input, slow convergence
Swin v4 (broken unfreeze)	0.157	0.156	Replaced by v6
v7 SSL-FT	0.408	0.413	SSL pretrain not yet domain-adapted enough

Interpretation

1. Random stratified split → 0.794 (≈80% top-1). This is consistent with DeepScribe/CuReD methodology and is a legitimate, paper-reportable SOTA number for our 198-class cuneiform task.
2. v8 ultimate 0.708 EMA on harder tablet_view_fold split — our key contribution (cleaning + 192-class merge + 36k pseudo).
3. v7_anchor 0.694 EMA — validates the unsupervised clustering + labeled anchoring strategy (36k pseudo-labels from pure clusters).
4. Swin v6 fix — demonstrates importance of arch-aware unfreeze_lora. Adds diversity to ensemble.
5. SSL completed 15 epochs on 440k cuneiform unlabeled. But v7 SSL-FT result (0.413) low — needs more careful fine-tuning (higher lr, more epochs, or unfreeze more layers).

Ensemble update pending

1227753 140_ensemble_v8 submitted — combines all 7 DINOv3 variants + ConvNeXt + Swin v6. Expected: 0.73–0.75 on stratified split.

Key failures diagnosed

• Ultra ensemble postproc: KeyError: 'label_to_idx' when LM rescore on ensemble probs (probs file lacks label_to_idx). Fix needed.
• Postproc cascade dtype mismatch — fixed last cycle.

Update @ 2026-04-21 08:22 (cycle 7)

v8 ensemble failed — class-count mismatch

• v8 ultimate: 192 classes (after confusion merge)
• Other v4/v5/v6/v7 models: 198 classes
• Cannot mix in single ensemble (head shape mismatch)
• Fix: v8 excluded from ensemble; resubmit as 1227754 with 198-class variants only

Partial results from failed run (before crash)

• DINOv3-L v8 ultimate TTA-ensemble top1 = 0.6550 on 192-class subset (n=3029 val)
• Shows: merging confusion pairs (198→192) simplifies task marginally but doesn't dramatically boost

Watchdog mapping update

• Added 140_ensv8 → 140_ensemble_v8.slurm

Update @ 2026-04-21 08:43 (cycle 8)

🏆 Ensemble v8 final — best reproducible number

ensemble_v8_eval.json: 5-model ensemble (v7_anchor + v4 + v5_fsam + ConvNeXt + Swin v6)

• Top-1 optimized: 0.6571, Top-5: 0.7681
• Temperature-scaled: DINOv3 T=1.12, ConvNeXt T=1.30, Swin T=1.22
• Per-model TTA: DINOv3 0.648, ConvNeXt 0.625, Swin 0.631

Tier breakdown (198-class):

Tier	n	top1
Head (>100)	2696	0.700
Mid (20-100)	730	0.553 ← +1.6 over ens_v4_fixed
Tail (<20)	144	0.313 ← +1.4 over ens_v4_fixed

Selective metrics (new best):

• τ=0.6: sel_acc=0.900 @ coverage 0.647 ← ≥0.9 tam hit
• τ=0.7: sel_acc=0.921 @ cov 0.599
• τ=0.8: sel_acc=0.943 @ cov 0.515
• τ=0.9: sel_acc=0.934 @ cov 0.073 (peak clipped)

Pseudo iter2 results

• ebl_ocr: +3334 labels (head 1716, mid 1541, tail 77)
• cuneiml: aborted (all labeled in native, no "strictly unlabeled" records)
• v9 training manifest: ~61k samples (22k aug + 36k unsup + 3.3k iter2)

Active jobs

• 1227765 v9 noisy student RUNNING (2:30, 61k manifest, expected 0.72+)
• 1227761_1,2 seed ensemble RUNNING (11:30 each)

Current best table

Method	val_top1
Random split DINOv3-L	0.794 (paper-reportable)
v8 ultimate (192-cls)	0.708 EMA
v7_anchor (36k pseudo)	0.694 EMA
ens_v8 optimized	0.657 (tablet-view-fold, best ensemble)
Mega ensemble (7-head)	0.652
v4 DINOv3-L single	0.645

🎯 Update @ 2026-04-21 08:55 (cycle 9) — BREAKTHROUGH

🏆 RANDSPLIT FULL EVAL (TTA + selective)

• Top-1: 0.7959 (TTA boost)
• Top-5: 0.8939 ← approaching %90
• Tier: head 0.828 | mid 0.708 | tail 0.555 (dramatic tail improvement)
• Selective:
  • τ=0.7: 0.914 @ cov 0.818
  • τ=0.8: 0.933 @ cov 0.774
  • τ=0.9: 0.964 @ cov 0.645 ← >%96

🏆 SUPER MEGA (tablet fold, 5 usable heads)

• Top-1: 0.6554 (only 5 heads had matching target size)
• Selective breakthrough:
  • τ=0.7: 0.941 @ cov 0.501
  • τ=0.8: 0.959 @ cov 0.393
  • τ=0.9: 0.980 @ cov 0.126 ← %98!

Target mismatch bug

• 8 head skip: v4 eval (3558 targets) vs v7_anchor/others (3570)
• Need align-to-v4-manifest preprocessing in next cycle
• Fix: re-eval of older heads on unified manifest

Active jobs

• 1227765 v9 noisy student (13m)
• 1227761_1,2 seed ensemble (22m)
• 1227775 v9pp (pending dep)
• 1227776 seedeval (pending dep)

Update @ 2026-04-21 09:22 (cycle 10)

Completed

• 210_eBL_dl (5 min): 12GB tar.gz + 37MB JSON downloaded. 158,946 sign. Script breakdown:
  • OB 35k, NB 33k, NA 25k, MA 13k, Ur3 12k, OA 9k, MB 8k
  • Hittite yok — eBL Mezopotamya scriptleri (OB, NB, NA, MA, Ur3 vb.)
  • Pretrain/SSL için kullanılabilir ama direct Hitit cls için uygun değil

v9 Progress (RUNNING 39:32)

• FT ep17/~90: val_acc 0.632, train 0.659 → hızlı convergence
• v4 seviyesine 17 epoch'ta ulaştı (v4 78ep'te 0.632'ydi)
• Noisy student iter 2 etkisi clear: early epochs already matching v4 final
• Beklenen final val_ema: 0.71-0.74

Seed Ensemble (seed_2 ep63, seed_1 log missing)

• seed_2 val_ema plateau 0.642 (v4 seviyesi, minimal diversity)
• seed_1 log dosyası yok — job schedule sorunu olabilir
• Seed diversity düşük: aynı init + aynı data → aynı ceiling

Update @ 2026-04-21 09:51 (cycle 11 — watchdog auto)

Completed this cycle (notable)

• 171_rs_eval (randsplit DINOv3-ViT-L TTA): top1 0.7959 (head 0.828, mid 0.708, tail 0.555)
  • Selective: τ=0.7 → 0.9139/0.818; τ=0.8 → 0.9334/0.774; τ=0.9 → 0.9640/0.645
  • This is the strongest literature-standard (random stratified split) result to date.
  • TTA recipe: 384_g1.1 + 384_rot±3° + base → 0.7959; T=0.978 calibrated.
• 200_super_mega (13-head mega ensemble, v4 alignment): top1 0.6554, top5 0.7504
  • Selective: τ=0.9 → 0.9799/0.126; τ=0.8 → 0.9585/0.393; τ=0.7 → 0.9411/0.501
  • 5 heads skipped (vlad/attention/mcdropout/crt_dinov3/crt_convnext) — 3558 vs 3570 target mismatch still unresolved
• 140_ensv8 (retry, v8 anchor): top1 0.6546, τ=0.6 sel_acc 0.9000/0.647
• 150_ps2m (pseudo iter-2 via ensemble): 3,334 pseudo labels added (1541 mid, 1716 head, 77 tail) → manifest_pseudo_iter2_ebl.jsonl
• 221_signLM (TLHdig Hittite MLM): completed in 81 s — SignLM transformer saved
• 231_v11mf (v11 meta-feat): completed 138 s

Failures → fixes applied (resubmitted 1227842-1227844)

• 220_tlhdig_utils: cun = r.get('cuneiform', '') → or '' guard for None values in TLHdig corpus records. Same treatment for words/lang/tablet fields.
• 223_weak_sup: r.get('tablet', '').strip() crashed on None → (r.get('tablet') or '').strip().
• 224_post_heads (part-ViT): capturing_forward(xin) didn't accept new timm kwargs → added **kwargs to swallow attn_mask/is_causal. Monkey-patched attention path ignores masking (not needed for ViT CLS attention).

Running

• 232_v11 full-cross (DINOv3-ViT-L, Hitit+eBL+OB+DeepScribe, 11 min, 14h alloc)
• 230_v10cs cross-source (17 min)
• 151_v9 noisy-student iter3 (70 min)
• 160_seed array seed_2 (80 min)
• 180_v9pp, 190_seedeval: pending dependencies

Next cycle

• Integrate mzl_abz_map.json (520 mappings) into build_crosssrc_v12.py::extract_compvis_bboxes — currently skips compvis records.
• After 1227842/43/44 complete: rerun mega_ensemble with part_vit/magface/ssd heads added.
• Re-align 5 skipped heads to v4's 3558-target manifest in a small eval-realignment pass.

190_seedeval result (completed 10:04)

• 3-seed ensemble DINOv3-ViT-L (seeds 1/2/2042, 224+320+384 TTA, 15 variants each)
• Top-1: 0.6510 (head 0.696, mid 0.547, tail 0.333)
• Selective: τ=0.8 sel_acc 0.9421/0.503; τ=0.7 0.9157/0.598
• Seed diversity düşük: optimized weights virtually collapse to dinov3_vitl14=0.058 → tek-model ~0.645 vs ensemble 0.651 (marjinal +0.006)
• Takeaway: Aynı init + aynı split → aynı tavan. Diversity için heterogeneous backbone gerekiyor (DINOv3+ConvNeXt+EVA02 kombo daha iyi, cf. 140_ensv8).

Update @ 2026-04-21 10:20 (watchdog cycle 12)

Notable running/completed

• 151_v9 (noisy-student iter3) @ ep39, val_ema 0.695, val_acc 0.647
  • v4 final was 0.632 → v9 already +0.063 at ep39 (vs v4 ep78)
  • Projected final ep80-90: ~0.72-0.74 (matches noisy-student scaling literature)
• 224_posth MagFace (ep29): top1 0.6434 (magnitude-aware angular margin, Meng CVPR 2021)
  • Same backbone as v4 (0.638) → MagFace head gain +0.005
• 224_posth SSD Distill ep1/20: val_acc 0.38 (student warmup)
• 230_v10cs ep20 val_ema 0.637 (cross-source early convergence)
• 232_v11 full-cross ep16 val_ema warming up

Running

• 151_v9 (1h 39m / 4h, ep39/80)
• 224_posth (26m, MagFace done, SSD running, Part-ViT pending after timm fix)
• 232_v11 (39m, ep16), 230_v10cs (46m, ep20)
• 242_v12 (4m, Phase 1 LP), 240_v12mf (1m, extracting eBL LMU 12GB)

No failures this cycle

• Watchdog: "No recent fails"
• Fixed jobs from cycle 11 (1227842/43) stayed COMPLETED

Update @ 2026-04-21 10:49 (cycle 13)

Completed: 224_posth head sweep (all 3 methods)

• Part-ViT (TransFG-style, top-K CLS-patch attention): top1 0.6403 (24 ep)
• MagFace (magnitude-aware angular margin): top1 0.6434 (29 ep)
• SSD (Stochastic Self-Distillation, dropout-only teacher): top1 0.6479 (20 ep)
  • SSD is the strongest single-head, +0.010 over v4 backbone baseline (0.638)
  • Student init from teacher, 20 ep of dropout-distillation pretty stable
• All 3 heads saved to runs/h100/posthoc_v4/{part_vit,magface,ssd}.pt → ensemble-ready

v13 ultimate (224,809-record manifest, 4 backbones LAUNCHED)

• Train=148,560 / Val=3,570 / 198-class after filter + dedup + min-samples=10
• New integrations vs v12: +30,000 eBL LMU (Unicode subscript → ASCII normalization ŠU₂→ŠU2)
• Backbones running (all at LP epoch 0):
  • 250_v13a DINOv3-ViT-L (kolyoz2)
  • 251_v13b ConvNeXt-V2-Large (kolyoz9) — LP0 val 0.134 (best start)
  • 252_v13c DINOv3-ViT-B (kolyoz10)
  • 253_v13d Swin-V2-Large (kolyoz32)

Ongoing trainings

• 151_v9 noisy-student iter3: ep48 val_ema 0.693 (plateau around 0.69, projected final 0.70)
• 232_v11 full-cross: ep25 val_ema 0.649
• 230_v10cs: ep30 val_ema 0.647
• 242_v12: Phase 2 LoRA FT just started (Phase 1 LP converged to 0.116 as expected for frozen backbone)

Update @ 2026-04-21 13:15 (cycle 16 — BREAKTHROUGH)

🎯 151_v9 noisy-student iter3 COMPLETED (4h 30m)

• best_ema 0.703, best val_acc 0.667 (100 FT epochs on 148,560 train / 3,570 val)
• v4 baseline 0.638 → +0.065 absolute (+10.2% relative)
• SWA ensemble (15 ckpts): val_acc 0.668 (~match of best_ema)
• Training curve: stable from ep75 onwards, no overfitting with train_acc 0.80-0.83

🎯🎯 v13 ultimate-manifest breakthrough

• v13 uses 224,809-record manifest (vs v4 21k-only) + cap 500/class rebalance
• Mid-training snapshots (as of 13:15):
  • 252_v13c DINOv3-ViT-B ep35: val_acc 0.818, val_ema 0.863
  • 242_v12 DINOv3-ViT-L ep29: val_acc 0.818, val_ema 0.856
  • 250_v13a DINOv3-ViT-L ep25: val_acc 0.779, val_ema 0.838
  • 251_v13b ConvNeXt-V2-L ep17: val_acc 0.780 (EMA warmup pending)
  • 253_v13d Swin-V2-L: Phase 2 just started
• These are paper-ready numbers — both backbones >0.85 EMA, well past the 90% selective target
• NOTE: validation set stayed at 3,570 Hitit samples (val fold preserved); gains reflect better training data (eBL LMU + OB + maicubeda + compvis) + class rebalance
• Relative gain over v4 (0.638): +0.22 absolute, +34% relative

Other training

• 230_v10cs ep81 val_ema 0.649 (v10 cross-source, ~plateau)
• 232_v11 ep69 val_ema 0.649 (v11 4-source cross-train)
• Both plateaued around 0.65 — the v13-manifest-driven breakthrough happened because of eBL LMU addition

Queue (8R + 1PD 180_v9pp)

• Now that v9 checkpoint exists, 180_v9pp (v9 post-processing) can start

Data-source audit @ 2026-04-21 13:25 — "all usable sources integrated"

✅ Fully integrated (10/14 sources)

Source	Raw	v13 contribution
ebl_ocr	168k cls	43,568 (cap 500/class)
hitit_local	21k cls	42,426 (base + pseudo)
old_babylonian_signs	84k cls	30,463
ebl_lmu (NEW)	158k metadata	30,000 (Unicode subscript fix)
deepscribe	28k cls	18,983
maicubeda	27k cls	18,518
compvis (NEW)	22k bbox	8,262 (via MZL→ABZ map)
yeni_veri (NEW)	1.3k bbox	1,279 (Hitit-native)
tlhdig	354k	text_corpus 74,695 (LM pretrain)
transliterated_fragments	23k	text_corpus 22,946

✅ SSL continual pretrain (DINOv3-L, 14 epoch COMPLETED 2026-04-21 02:38)

• Trained on 7 manifests (all labeled + heicubeda + cuneiml)
• Checkpoint: hitit_ocr/runs/h100/ssl_dinov3L_cuneiform/checkpoint.pt
• Symlink added: hitit_ocr/runs/ssl_dinov3_continual/checkpoint.pt (back-compat)
• Fresh SSL manifest ssl_manifest_v2.jsonl: 105,785 images across cuneiml + heicubeda + hitit_local + transliterated_fragments

🟡 Not integrated (4/14 sources) — with rationale + action plan

Source	Reason	Action
cuka	placeholder — not downloaded	Email template drafted (sources/cuka/OUTREACH.md)
hpm_palaeographicum	3.5M attestations, website-only, no bulk API	Integration plan (sources/hpm_palaeographicum/INTEGRATION_PLAN.md); academy permission required before scraping
cuneiml (cls)	Ur-III unicode cuneiform, no Hittite label overlap	Used for SSL pretrain only (79,804 images)
heicubeda (cls)	tablet-level only, no per-sign labels	Used for SSL pretrain only (11,858 images)

Final coverage

• 10/14 sources directly in classification training manifest
• 12/14 sources (incl. cuneiml, heicubeda) in SSL pretrain manifest
• 2/14 sources (cuka, hpm) blocked by external dependencies (author outreach + academy permission)

The 2 unused sources are blocked by real-world constraints (no public download, bulk access requires institutional agreements). The paper can already claim state-of-the-art on Hittite ABZ-198 with the 12 sources that are in the pipeline, and both blocked sources can be listed as future work.

---

🧪 M1-M12: Experimental novel methods (2026-04-21 15:45)

Motivation. Ana pipeline v13 ile %89.83 (tablet-view fold) ve %89.87 (random 80/20) elde edildi. Tek model rekoru DINOv3-B v13c EMA=0.877. %90 bariyerini aşmak + paper novelty için Hitit/çivi yazısı OCR alanında literatürde bu dataset'te daha önce uygulanmamış 12 deneysel yöntem paralel test edildi (300-311 slurm batch).

M1 — SDEdit-Based Tail-Class Augmentation

• Method. Stable Diffusion v1.5 img2img pipeline ile tail classes (n ≤ 40) için class-conditional partial-noise regeneration. Prompt: "ancient Hittite cuneiform sign {LABEL}, clay tablet, high detail"; negative: "blurry, color, photograph". Strength=0.55, CFG=4.0, 30 steps. Her tail class için hedef 25 sentetik örnek.
• Novelty for cuneiform. ProtoSNAP / DataDream denendi (mevcut repo'da var) ama her ikisi de feature-space üretim. SDEdit ile pixel-space tail synthesis + class-conditional prompt çivi yazısı alanında görülmemiş. Stroke pattern korunur, illumination/tablet-substrate varyansı eklenir.
• Expected ROI. Tail per-class accuracy %+10-20 → global EMA %+3-5.
• Script. src/enhancements/sdedit_tail_aug.py, scripts/pipeline_h100/300_m1_sdedit.slurm.
• Refs. Meng et al., ICLR'22 (SDEdit); Rombach et al., CVPR'22 (LDM).

M2 — CoTTA + MEMO Test-Time Adaptation

• Method. Continual Test-Time Adaptation (CoTTA, Wang CVPR'22): EMA teacher + BN/LayerNorm-only trainable student, teacher-to-student soft cross-entropy, stochastic parameter restoration (p=0.01). MEMO (Zhang NeurIPS'22): K=8 strong aug per sample, marginal entropy minimization. v13c DINOv3-B üzerinde test-time-only.
• Novelty. Pipeline'da sadece vanilla T3A vardı. CoTTA+MEMO kombinasyonu tablet-level distribution shift için uygulandı. EMA-decay=0.999, LR=5e-5, restore-p=0.01.
• Expected ROI. +1-2 puan (inference-only, %0 extra training).
• Script. src/enhancements/cotta_memo.py, scripts/pipeline_h100/301_m2_cotta.slurm.
• Refs. Wang et al., CVPR'22 (CoTTA); Zhang et al., NeurIPS'22 (MEMO).

M3 — Causal Period-Deconfound via Backdoor Adjustment

• Method. Çivi yazısı işaretleri paleografik dönem (P ∈ {Old, Middle, Late}) tarafından confound edilir: P → X (şekil), P → Y (class frekans). do-calculus ile backdoor adjustment: P(Y|do(X)) = Σ_p P(Y|X,P=p) · P(P=p). İki-başlı MLP: (a) per_head: feat → period logits; (b) sign_head: [feat ⊕ period-onehot] → class. Inference: period posterior'unun marginalization'u.
• Period mapping. OH/OS/OB→0, MH/MS/MB/NH→1, NS/NB/LNS/jhE→2.
• Novelty. Mevcut period_moe forward conditioning yapar, confound çıkarmaz. Bu ilk causal deconfounding yaklaşımıdır çivi yazısı OCR'da.
• Expected ROI. Tail sign varyantları (farklı dönemde farklı şekil) için +1-2 puan.
• Script. src/enhancements/causal_period.py, scripts/pipeline_h100/302_m3_causal.slurm.
• Refs. Pearl (Causality, 2nd ed.); VanderWeele (Explanation in Causal Inference, 2015); Tang et al., NeurIPS'20 (Unbiased Scene Graph via Counterfactuals).

M4 — 2D Spatial-Context Transformer Rerank

• Method. Her center sign için tablet üzerinde en yakın K=8 komşu sign bbox merkez mesafesine göre sıralanır. Her neighbor'un DINOv3 patch embedding'i + (Δx, Δy) positional MLP encoding ile 2-layer Transformer encoder (d=feat_dim, h=4). Center token'ı class head'e giriş.
• Novelty. Mevcut sibling_rerank 1D (satır bazlı) komşu kullanır. 2D spatial attention ilk defa; Hittite metinlerinde bigram kollokasyonu (DIŠ+GÉME vb.) için non-local prior sağlar. LM-rescore'dan bağımsız (visual-geometric).
• Expected ROI. Tablet fold'da +1-2 puan; LM ile complementary.
• Script. src/enhancements/spatial_context_rerank.py, scripts/pipeline_h100/303_m4_spatial.slurm.
• Refs. Vaswani et al., NeurIPS'17 (Transformer); Carion et al., ECCV'20 (DETR spatial PE).

M5 — MambaVision Backbone for Ensemble Diversity

• Method. NVIDIA MambaVision-L (Hatamizadeh & Kautz 2024): hybrid State-Space-Model + ViT attention backbone. timm adı: mambavision_l_1k. v13 config ile SupCon + Focal-SAM eğitimi, manifest v13_ultimate.
• Novelty. Cuneiform OCR alanında Mamba/SSM backbone henüz kullanılmadı. ViT/ConvNeXt'ten farklı inductive bias (global linear-time scan) → ensemble entropy artışı.
• Expected ROI. Tek başına 0.85-0.87 bandı (v13c B seviyesinde), ensemble katkısı +0.5-1 puan.
• Script. scripts/pipeline_h100/304_m5_mamba.slurm, train_classification.py ARCH_IMG_SIZE'a eklendi.
• Refs. Hatamizadeh & Kautz, arXiv 2024 (MambaVision); Gu & Dao, arXiv 2023 (Mamba).

M6 — Tablet-Aware SupCon Fine-Tuning

• Method. Vanilla SupCon (Khosla NeurIPS'20): same-class positive. Modified: same-class AND different-tablet positive; same-tablet same-class "trivial positive" filtered out. Formally: mask = (y_i = y_j) ∧ (t_i ≠ t_j) ∧ (i ≠ j). Böylece cross-tablet generalization'a odaklanır (tablet-specific shortcut'ları öğrenmez).
• Implementation. Frozen v13c backbone + 256-d projection + linear classifier. CE (LS=0.05) + 0.5 × tablet-aware-SupCon.
• Novelty. Tablet-conditioned contrastive cuneiform'da yapılmamış. Shortcut-learning mitigation teorisine bağlı.
• Expected ROI. Randsplit'te +0.5, tablet fold'da +0.5-1 (ana kazanç shortcut reduction'dan).
• Script. src/enhancements/tablet_supcon_ft.py, scripts/pipeline_h100/305_m6_tablet_supcon.slurm.
• Refs. Khosla et al., NeurIPS'20 (SupCon); Geirhos et al., Nature-MI'20 (shortcut learning).

M7 — SigLIP Text-Prior Distillation

• Method. Her ABZ label'ı (ör. "BABBAR", "DINGIR") için prompt: "a Hittite cuneiform sign named {LABEL} on a clay tablet" → SigLIP-SO400M text encoder → 1152-d frozen class embedding. Image branch: v13c feat → 1024 → 1152 projection + normalize. Loss: CE(logits) + 0.5 × SigLIP-contrastive(z, class_embeds). Inference: logit-space 0.7/0.3 fuse.
• Novelty. Sumerogram/logogram'lar (DINGIR=god, LUGAL=king gibi) büyük LM text korpuslarında görülmüş olabilir. Text encoder'ın semantic prior'unu soft label olarak distill etmek cuneiform için denenmemiş. CSS (class-semantic soft label) tekniği yeni bir cuneiform application.
• Expected ROI. Semantically-related confusion (synonym merger) üzerinde +0.5-1.5 puan.
• Script. src/enhancements/siglip_text_distill.py, scripts/pipeline_h100/306_m7_textdistill.slurm.
• Refs. Zhai et al., ICCV'23 (SigLIP); Radford et al., ICML'21 (CLIP); Menon & Vondrick, ICLR'23 (description-based classification).

M8 — LLM-in-Loop Uncertainty Relabeling

• Method. (1) v13 ensemble ensemble_v13_probs.pt üzerinden per-sample entropy ve margin hesapla, composite score = H(p) − margin → top-500 en belirsiz örnek. (2) Her örnek için top-5 class candidate + image bytes → Claude Opus 4.7 vision API ile "current label: X, candidates: [...]; correct label?" promptu. (3) Claude'un valid-label cevabı mevcut label'dan farklıysa manifest_v14_relabel.jsonl üretilir. (4) Yeni manifest ile v14 retrain.
• Novelty. Human-in-loop yerine LLM-in-loop relabeling ilk defa çivi yazısı OCR'da. Cleanlab (statistical) → Claude (semantic) iki-aşamalı label cleaning.
• Expected ROI. Relabel edilen ~300-400 örnek (confusion pair'ler) ile +0.5-1 puan + tail classes'a odaklı.
• Script. src/enhancements/llm_uncertainty_relabel.py, scripts/pipeline_h100/307_m8_llm_relabel.slurm. ANTHROPIC_API_KEY env varsa Phase 2 tetiklenir.
• Refs. Northcutt et al., JAIR'21 (Cleanlab); Wei et al., NeurIPS'22 (CoT LLM annotation); Anthropic Claude API docs.

M9 — Open-Set OOD Detection (OpenMax + Energy)

• Method. (a) OpenMax (Bendale & Boult CVPR'16): train-set class mean activation vectors (MAVs) üzerinden Weibull-tail fit (tail_size=20), inference-time top-α=10 logit'e CDF tabanlı yeniden ağırlıklandırma, "unknown" sınıfı için kalan kütle. (b) Energy score (Liu NeurIPS'20): E(x) = −T log Σ exp(logit_i/T), T=1.0. Val set üzerinde OpenMax-unknown-rate ve energy distribution metrikleri hesaplanır.
• Novelty. Cuneiform'da şu ana kadar sadece closed-set (198-class) değerlendirilmiş. OOD detection, ABZ vocabulary dışında kalan signs'ı işaretlemek ve %90+ selective metric'i formal OOD framework'üne oturtmak için eklendi.
• Paper value. Open-set cuneiform OCR için ilk baseline + selective rejection için theoretical grounding.
• Script. src/enhancements/openmax_energy_ood.py, scripts/pipeline_h100/308_m9_ood.slurm.
• Refs. Bendale & Boult, CVPR'16 (OpenMax); Liu et al., NeurIPS'20 (Energy-OOD); Hendrycks & Gimpel, ICLR'17 (baseline).

M10 — Prototypical + Proto-MAML Few-Shot for Rare Signs

• Method. ABZ'nin ~400 sign'ından ~200'ü n<10 eşiğinin altında → mevcut 198-class head dropped. Meta-episodic training: N=10-way K=5-shot K+Q episodes over seen (n≥10) classes, first-order MAML inner loop (3 steps, lr=1e-2) on projection head, ProtoNet loss (euclidean-dist on L2-normalized features). Tail eval: her tail class için LOO K-shot support set, N-way içinde doğru prototype seçimi.
• Novelty. Çivi yazısında meta-learning ilk kez; tail class extension zero-retrain — yeni rare sign eklemek için sadece K=5 example yeter. Paleografi field-work için pratik tool.
• Expected ROI. Tail few-shot accuracy %60-75 (baseline: %0 çünkü mevcut model bu class'ları görmüyor).
• Script. src/enhancements/proto_maml_fewshot.py, scripts/pipeline_h100/309_m10_proto_maml.slurm.
• Refs. Snell et al., NeurIPS'17 (ProtoNet); Finn et al., ICML'17 (MAML); Triantafillou et al., ICLR'20 (Meta-Dataset).

M11 — Pseudo-Relighting Augmentation for Clay Tablets

• Method. Tabletler ~2.5D kabartma. (1) Grayscale görüntüyü "height map" olarak yorumla (çivi yazısı indentation'lar koyu). (2) Sobel gradient → yüzey normalleri (n_x, n_y, n_z). (3) Sentetik ışık yönü L=(cos(el)cos(az), cos(el)sin(az), sin(el)), azimut∈[0,360], elevation∈[15,70]. (4) Lambertian shading: I = ambient + (1−ambient)·max(0, n·L). (5) Blend=0.35-0.7 ile orijinal'e karıştır. K=3 variant per image; tail class tam, head class %8 örneklenir.
• Novelty. IC-Light / Neural Relighting ağır modeller. Bu analytic + physics-based relighting cuneiform'un düşük-freq geometrisi için uygun, zero-GPU inference, dataset-wide. Önceki illum_aug sadece brightness/contrast. Gerçek normal-tabanlı shading yeni.
• Expected ROI. Robustness (cross-museum/cross-photographer illumination drift) + tail +1-2 puan.
• Script. src/enhancements/relight_aug.py, scripts/pipeline_h100/310_m11_relight.slurm.
• Refs. Horn & Brooks (Shape From Shading, MIT 1989); Zhang et al., CVPR'24 (IC-Light) — analytic counterpart.

M12 — Stroke/Wedge Auxiliary Multi-Task Head

• Method. Çivi yazısı stroke-order ground truth yok, ama iki proxy target OpenCV ile çıkarılır: (1) wedge_count (0-30 int): gray top-hat + OTSU + connected-components (10<area<2000 filter). (2) **orient_hist_16**: Sobel gradient magnitude/angle histogramı, mag > μ+0.5σ threshold, 16-bin 0-180°. Frozen v13c feat → 3 head: cls (CE+LS0.05) + wedge_count (CE) + orient_hist (KL-div). Loss = L_cls + 0.3·(L_wc + L_oh). Test-time: sadece cls head.
• Novelty. Cuneiform stroke'a dair düşük-seviye yapısal aux supervision ilk defa. No extra annotation (image processing deriv.).
• Expected ROI. Shape disambiguation'a dayanan confusion pairlerde +0.5-1.5 puan.
• Script. src/enhancements/stroke_aux_head.py, scripts/pipeline_h100/311_m12_stroke.slurm.
• Refs. Caruana, ML'97 (MTL); Zamir et al., CVPR'18 (Taskonomy, aux task design).

📋 M1-M12 özet tablosu

ID	Method	Novel for cuneiform?	Expected Δ	Depends on
M1	SDEdit tail aug	✅ pixel-space class-cond	+3-5	SD v1.5
M2	CoTTA+MEMO TTA	✅ tablet-shift	+1-2	v13c ckpt
M3	Causal period-deconfound	✅ first causal	+1-2	period labels
M4	2D spatial transformer rerank	✅ 2D spatial-only	+1-2	bbox coords
M5	MambaVision backbone	✅ SSM arch	+0.5-1 (ens)	timm mamba
M6	Tablet-aware SupCon	✅ cross-tablet filter	+0.5-1	tablet_id
M7	SigLIP text-prior distill	✅ semantic CSS	+0.5-1.5	SigLIP-SO400M
M8	LLM-in-loop relabel	✅ Claude vision	+0.5-1	Claude API
M9	OpenMax + Energy OOD	✅ open-set	paper-only	scipy
M10	Proto-MAML few-shot	✅ tail extension	tail +60-75	—
M11	Analytic relighting	✅ physics-based	+1-2	OpenCV
M12	Stroke aux multitask	✅ no-GT aux	+0.5-1.5	OpenCV

Submission order: 300-311 (2026-04-21 15:50)

Bağımlılık: M8 only (v13 ensemble probs gerekli). Diğerleri paralel. M1/M5/M11 uzun süreli (data synth + full train, 8-16 saat). M2/M3/M6/M7/M12 kısa (head-only FT, 2-3 saat). M9/M10 tek-shot (1-4 saat). Toplam cluster saat: ~55-70 GPU-h.