TRACCERR/Sumtables-Cuniform-Small-Fable5-Remaster-v2

Sumtablets Cuneiform — Small (Fable5 Remaster v2)

This is the exact multi-task training/eval/test pack used to fine-tune Qwen3-VL-4B-Instruct to visually read, transliterate, and translate Sumerian cuneiform tablets from photographs (see the companion model repo, linked below). It is a fixed, reproducible subset drawn from the full master dataset:

Master dataset (455,506 records, 70.5 GB, all phases, full documentation): TRACCERR/Sumtables-Cuneiform-Full-Fable5-Remaster

If you need the full corpus, more tasks, more tablets, or the underlying pipeline/phase documentation, go there. This "Small" repo exists specifically so the exact data behind a specific trained model checkpoint is preserved, versioned, and reproducible on its own — not as a replacement for the master dataset.

What's in this repo

File	Split	Rows	Notes
data/train-*.parquet	train	43,650	Sampled from the master dataset per the task shares below. T5_sign_grounding rows have been removed from this public copy (see Licensing)
data/validation-*.parquet	validation	720	Used for in-training eval / early stopping / best-checkpoint selection
data/test-*.parquet	test	24,446	Full held-out test split from the master dataset (unsampled) — used for all post-training quantitative/qualitative evaluation
reports/starter_train_v2_report.json	—	—	Exact per-task sampling report for the train split (targets, selected counts, tier composition, refusal counts)
reports/eval_pack_v2_report.json	—	—	Same, for the validation split

Each row is one training example: a system prompt, a user turn (instruction

• optionally an embedded tablet photograph or lineart image), and the target assistant response, in the same conversations JSON format used by the training pipeline (OpenAI-style chat turns). Rows carry task, tablet, split, tier, provenance, image (embedded JPEG bytes, where the task is vision-based), and conversations columns.

Task composition (train split)

The train split was sampled from the master dataset using fixed per-task shares, then deduplicated by tablet within each task:

Task	Share	Rows selected	What it teaches
T1_surface_ocr	25%	12,125	Read one tablet surface (photo) → line-by-line transliteration
T3_lineart_reading	15%	7,275	Read one tablet surface from a hand-drawn lineart scan → transliteration
T10_abstention	15%	7,275	Recognize genuinely illegible/damaged images and correctly decline (<ILLEGIBLE_IMAGE>) instead of guessing
T2_full_tablet	8%	3,880	Read an entire multi-surface tablet (obverse/reverse/edges) in one pass
T6_signs_to_translit	7% (target)	3,395	Text-only: cuneiform sign sequence → transliteration
T8_translation	7% (target)	3,395	Transliteration → English translation
T4_photo_lineart_pair	5%	2,425	Photo + matching lineart shown together → transliteration
T11_metadata	5%	2,425	Photo → structured metadata (period, genre, provenience)
T7_translit_to_signs	3% (target)	1,455	Text-only: transliteration → cuneiform sign sequence
T5_sign_grounding	10% (target)	excluded from this repo	Sign-level bounding-box grounding (external eBL corpus) — see Licensing

T10_abstention specifically contains a mix of genuine damaged/illegible tablets (real refusal targets, tier C, 2,888 of the 7,275) and legible tablets phrased as the same task type (tier A, 4,387) — this balance is intentional and matches the exact data the linked model was trained on. The refusals field in starter_train_v2_report.json is a verified true count (confirmed by direct inspection of the assistant-turn content, not merely the build report — see the model card for why that distinction matters).

The validation and test splits use the same task set, at their own natural sampling (validation: hand-picked small eval pack, 720 rows; test: the master dataset's full, untouched test split, 24,446 rows, entirely disjoint from train/validation at the tablet level — see the master dataset card for the leakage-prevention methodology).

Licensing — by component, not blanket

This dataset mixes content with different upstream licenses. Matching the master dataset's approach, license is declared per component rather than as a single blanket term:

• Text (transliterations, translations, metadata), upstream SumTablets corpus: CC BY 4.0
• Tablet photographs, upstream: Apache 2.0
• CDLI-sourced ATF content: governed by CDLI's own attribution terms
• Pipeline / task-construction code (not included in this repo, see the master dataset repo): MIT

T5_sign_grounding has been fully excluded from this repo. That task draws on external sign-bounding-box imagery from the eBL (electronic Babylonian Library) corpus, which has no stated redistribution license upstream. The master dataset documents this explicitly and keeps that task in a separate file specifically so it can be omitted from any public redistribution — this repo follows that same rule. If you need T5 data for research use, it remains available train-side-only in the master dataset repo under its own documented terms; it is intentionally not present here.

Provenance and reproducibility

This is not independently curated — it's a deterministic sample of the master dataset, produced by the same pipeline's phase2_tasks.py starter command against a fixed seed and an explicit scaleup_shares.yaml mixture. Anyone with access to the master dataset and pipeline code can regenerate this exact file set. It is published here as a fixed, versioned artifact tied to a specific trained model release, not as the primary distribution channel for the underlying corpus.

Related

• Master dataset (full corpus + pipeline docs): TRACCERR/Sumtables-Cuneiform-Full-Fable5-Remaster
• Model trained on this data: TRACCERR/Qwen3-VL-4B-Instruct-Cuneiform-Fable5-Remaster-V2-bnb-4bit (private)