Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Tasvir Bankası is provided for research, education, reproducibility review, and non-commercial evaluation only. Commercial use, resale, paid text-generation products, commercial model training, commercial derivative dataset production, and author-style cloning or imitation services are not permitted without explicit written permission from the author.
Log in or Sign Up to review the conditions and access this dataset content.
Tasvir Bankası: Turkish Literary Scene-State-Description Dataset
A gated non-commercial research dataset for Turkish literary scene, state, and description annotation.
Tasvir Bankası is a Turkish literary scene-state-description dataset and reproducible annotation pipeline release prepared by Furkan Yaşar. It provides structured JSONL records derived from rights-reviewed public-domain candidate Turkish prose, including segmentation, dialogue, tense, scene-boundary, state, and descriptive evidence fields.
This release is intended for research, education, non-commercial evaluation, citation, review, and reproducibility. It is not a style-cloning resource, author-imitation corpus, prose-style cloning dataset, or benchmark-grade full gold corpus.
DOI and publication record
The publication manifest archive for this release is registered on Zenodo.
Zenodo DOI:
https://doi.org/10.5281/zenodo.20579958
Recommended dataset citation:
Yaşar, F. (2026). Tasvir Bankası v0.3.1: A Turkish Literary Scene-State-Description Dataset and Annotation Pipeline. Zenodo. https://doi.org/10.5281/zenodo.20579958
Version relationship
Pipeline and dataset versions are tracked separately.
- Public-safe GitHub pipeline package:
v0.1.0-public-preview - Full gated dataset package:
v0.3.1
The v0.3.1 dataset is not a new pipeline run. It is a publication-ready release derived from the v0.3 curated package: filename normalization, source-inventory alignment, source-edition metadata recording, label-level gold-eval clarification, limitation-note strengthening, and regenerated manifest/checksum reports.
Dataset structure
data/train/*.jsonl
data/gold_eval/curated_gold_eval_v0.3.jsonl
data/curation_ledger/human_review_curation_ledger_v0.3.jsonl
docs/
reports/
Splits
- Train: pipeline-generated silver/silver-plus records.
- Curated gold eval v0.3: curator-reviewed, label-level evaluation seed.
- Curation ledger v0.3: former manual-review decisions preserved as resolved curation records; pending count is zero.
Current release metrics
train_record_count: 2172
curated_gold_eval_v0.3_count: 92
curation_ledger_v0.3_count: 98
pending_queue_count: 0
Annotation layers
The dataset includes structured fields related to:
- scene segmentation and scene-source metadata,
- state identity profile,
- descriptive profile family,
- enriched and filtered descriptive evidence,
- dialogue signals,
- tense and morphology features,
- selection and validation metadata.
Representative field families include:
scene_segments
scene_source
scene_edge
state_identity_profile
descriptive_profile
descriptive_profile_enriched
descriptive_profile_enriched_filtered
dialogue
tense
morphology
selection_metadata
Two-regime scene policy
brief_summary is not globally mandatory for every record.
The release follows a two-regime validation policy:
- Event-bearing scenes are expected to include
brief_summary. - Non-event, state, emotion, description, or structural-fallback segments may omit
brief_summarywhen accepted alternate evidence fields are present. - Summary-less records without accepted alternate evidence are invalid.
This policy prevents non-event descriptive or state-bearing literary units from being incorrectly rejected only because they lack conventional scene summaries.
Descriptive extraction policy
Descriptive extraction is regression-tested and cleaned against known lexical false-positive families. Ambiguous edge cases are routed to the curation ledger.
The descr_metaphor field should be read as evidence for metaphor, simile, analogy, comparison, or figurative-marker behavior. It should not be interpreted as a complete conceptual-metaphor analysis layer.
Curation policy
Records requiring semantic judgment are stored in:
data/curation_ledger/human_review_curation_ledger_v0.3.jsonl
The curation ledger preserves curator decisions such as:
resolved_true_metaphor_marker
resolved_removed_false_label
approved_for_gold_eval
The current release has:
pending = 0
Gold-eval policy
The file:
data/gold_eval/curated_gold_eval_v0.3.jsonl
contains 92 curated records.
This file is a label-level evaluation seed, not a full-record benchmark-grade gold corpus. Most records are marked with:
"gold_scope": "single_label_verified"
This means that the listed curation.verified_dimensions are curator-reviewed. It does not mean that every field of the entire record is verified as a complete benchmark-grade annotation.
Source texts and rights status
Source texts are documented as rights-reviewed public-domain candidates under the Turkish death+70 rule.
This release does not constitute a legal determination. Users are responsible for verifying source and jurisdictional status before reuse.
Modern publisher material, editor notes, introductions, afterwords, footnotes, simplified rewrites, cover matter, and other publisher-specific paratext are excluded from the intended source boundary.
Relevant documentation:
docs/SOURCE_TEXT_INVENTORY.md
docs/SOURCE_RIGHTS_REVIEW_TABLE.md
docs/ZENODO_LICENSE_NOTE.md
Source corpus rationale
The initial Tasvir Bankası release focuses on rights-reviewed public-domain candidate prose by Sabahattin Ali and Sait Faik Abasıyanık.
This source choice is methodological rather than incidental. Both authors fall within a plausible death-plus-70 public-domain candidate window under the Turkish copyright framework, although this release does not constitute a legal determination. Their prose is also relatively close to modern Turkish when compared with older Ottoman-influenced public-domain prose, making it more suitable for a first Turkish literary NLP resource focused on scene, state, dialogue, tense/morphology, and descriptive evidence.
The two-author focus provides genre and narrative-function coverage across novelistic prose, short fiction, psychological narration, dialogue, urban observation, atmosphere, and description. The release therefore does not attempt to process arbitrary public-domain Turkish prose. It begins with a controlled modern-Turkish literary core before extending to more historically or linguistically complex source pools in future versions.
This design choice also clarifies the first-stage scope of the project. Older public-domain authors with heavier Ottoman Turkish vocabulary and stronger diachronic language features remain valuable for historical Turkish or Ottoman Turkish NLP. However, they may introduce avoidable historical-language noise into an initial scene-state-description dataset intended to model modern Turkish literary prose. Tasvir Bankası therefore prioritizes a linguistically accessible, canonically significant, and computationally tractable source base for its first release.
License and access
Tasvir Bankası v0.3.1 is released as a gated, non-commercial research dataset.
- Dataset records, derived annotations, schemas, reports, documentation, and sample data are licensed under CC BY-NC 4.0, unless otherwise noted.
- Companion pipeline code is licensed under the PolyForm Noncommercial License 1.0.0.
- Access is gated to ensure that users acknowledge the research and non-commercial terms of use.
See:
LICENSE.md
CODE_LICENSE.md
DATA_LICENSE.md
TERMS_OF_USE.md
Permitted use
This dataset may be used for:
- academic research,
- education,
- non-commercial evaluation,
- reproducibility studies,
- Turkish literary NLP research,
- digital humanities research,
- computational narratology research.
Prohibited use without written permission
The following uses are not permitted without explicit written permission from the author:
- commercial model training,
- resale,
- paid text-generation products,
- paid API or SaaS integration,
- commercial derivative dataset production,
- redistribution as part of a commercial dataset package,
- author-style cloning or imitation services.
Required acknowledgement for gated access
By requesting access, users acknowledge that this dataset is provided for research, education, reproducibility review, and non-commercial evaluation only.
Suggested gated access acknowledgement:
I agree to use Tasvir Bankası only for research, education, reproducibility review, or non-commercial evaluation. I will not use it for commercial model training, paid text-generation products, resale, commercial dataset generation, or author-style cloning without explicit written permission from the author.
Known limitations
See:
docs/KNOWN_LIMITATIONS.md
Current limitations include:
- the dataset is a carefully engineered pilot resource rather than a large-scale benchmark corpus,
- the gold-eval subset is label-level, not full-record benchmark-grade annotation,
- source-rights review is documented but does not constitute legal advice,
- Turkish-specific morphology, tense, lexical matching, and literary assumptions would need adaptation before transfer to other languages.
Related resources
Zenodo publication record:
https://doi.org/10.5281/zenodo.20579958
GitHub public-safe annotation pipeline package:
https://github.com/Kon-tiki-ship/tasvir-bankasi-turkish-literary-annotation-pipeline
Citation
Creator: Furkan Yaşar
Recommended citation:
Yaşar, F. (2026). Tasvir Bankası v0.3.1: A Turkish Literary Scene-State-Description Dataset and Annotation Pipeline. Zenodo. https://doi.org/10.5281/zenodo.20579958
Use CITATION.cff for machine-readable citation metadata.
Contact and permission
For commercial licensing, redistribution permission, collaboration, or research questions, contact the author through the repository or associated publication record.
- Downloads last month
- 38