anhalu/vncompress-vi-dataset

VNCompress-VI: Vietnamese Context Compression Dataset

Training and evaluation data for context compression / extractive summarization on Vietnamese text, built for the VNCompress project (tone- and morphology-aware context compression for Vietnamese LLMs).

Two configs:

• records (train / validation, document-level 90/10 split) — the full record stream: raw corpus paragraphs, VCC-Bench samples, and teacher-distilled (context, query, compressed_text) triples with compression-quality metadata (target_tokens, realized_tokens, extractive_ratio, numbers_preserved, ...). This is what a compressor or relevance-probe model should train on.
• vcc_bench (train / validation / validation_synthetic) — the same content reshaped into the benchmark format the project's benchmark.py harness consumes directly.

⚠️ Read before reporting numbers on vcc_bench.validation

vcc_bench.validation (vcc_bench_eval.json) has 132 samples, but only ~6 of them have a reference_answer that is not a near-verbatim copy of context. On the other ~126, every compression-quality metric (ROUGE-L, BERTScore, EM) degrades to "how much raw text overlap survived compression" rather than measuring answer quality — so any method that keeps more tokens will look artificially better on those samples. This is a known, documented gap in the source project (see provenance/verification_report.json and the upstream 2026-09-06_data-pipeline-run.md report), not a corruption introduced by this upload. Do not treat scores over the full 132 as a valid benchmark. validation_synthetic (teacher-generated) must never be pooled with validation (independent) — scoring a model on its own teacher's output inflates numbers in the model's favor.

compression_ratio in records.train/records.validation metadata is the requested ratio, not the realized one — at a requested 2x the teacher distillation run in this dataset actually achieved ~7.3x. Read realized_tokens / target_tokens if you need the true ratio.

Construction

raw sources -> normalize -> verify -> document-level stratified split -> teacher distillation (GLM-5.2) -> filter. Split is a pure function of (document content hash, seed=42) — not random — so document-level train/eval separation is reproducible and leak-checked (provenance/split_manifest.json records the full eval document-key list and a leakage assertion result: CLEAN).

total records (records config)	141,294
train / validation	127,189 / 14,105
documents	3,412 (2,796 train / 616 validation)
accepted teacher-distilled compressions	57,529
teacher model	GLM-5.2 (OpenAI-compatible endpoint)

See provenance/ in this repo for the full audit trail: split_manifest.json, records_meta.json, verification_report.json, PROVENANCE.md, CHECKSUMS.json.

Licenses by source

This dataset unions text from multiple sources with different licenses. If you plan to redistribute a filtered subset, filter by source /metadata and re-check the applicable license for that subset rather than assuming the top-level cc-by-sa-4.0 tag covers everything permissively:

source field	License	Notes
uvw-2026, wikipedia	CC BY-SA 4.0	Vietnamese Wikipedia (via undertheseanlp/UVW-2026 and direct Wikimedia API fetch)
vcc_bench (legal-text samples)	Public domain	Vietnamese state legal documents
vcc_bench (synthetic templates: agent/cross-lingual/multi-turn/needle)	MIT	Hand-authored
vietnamese-poetry	MIT (content), but sourced from bigscience-data/roots_vi_vietnamese_poetry, a gated dataset requiring acceptance of the BigScience Ethical Charter	Included here ungated at the source-project maintainer's discretion; if you are re-hosting or redistributing further, review the BigScience Ethical Charter yourself before doing so
teacher-synth	MIT (project-original prompts/labels)	Generated by GLM-5.2 via an internal FPT Cloud endpoint; distillation outputs, not verbatim third-party text

Known limitations

• Only ~6/132 vcc_bench.validation samples are genuinely independent QA (see warning above).
• numbers_preserved is 0.91–0.94 across accepted compressions: the teacher drops numbers when the query doesn't ask about them; it does not invent them (number_altered fired 2/57,529 times).
• No supervised student model has been trained on compressed_text yet as of this dataset snapshot — it is offered as training data, not as proof a downstream model already benefits from it.

Citation / provenance

Built by the VNCompress project. See provenance/PROVENANCE.md for the full source-by-source build log and regeneration commands.