README.md

metadata

language:
  - lv
license: other
license_name: research-use-mixed-upstream-data
library_name: transformers
pipeline_tag: summarization
base_model: facebook/mbart-large-50-many-to-many-mmt
base_model_relation: finetune
tags:
  - summarization
  - mbart50
  - low-resource
  - research
  - rahvusarhiiv
datasets:
  - dennlinger/eur-lex-sum
  - FMISummarization/FMI_Summarization
widget:
  - text: >-
      Latvijas pētnieki paziņoja par jaunu projektu, kura mērķis ir uzlabot
      digitālo arhīvu pieejamību. Projekts paredz sadarboties ar bibliotēkām,
      universitātēm un valsts institūcijām.
    example_title: Latvian example
model-index:
  - name: lv_summariser revision 2
    results:
      - task:
          type: summarization
          name: Summarization
        dataset:
          name: Strict Wiki4 (wiki4_lv)
          type: wiki4_lv
        metrics:
          - name: ROUGE-L F1
            type: rougeL
            value: 0.313275
          - name: Reference content-token coverage
            type: reference_content_token_coverage
            value: 0.290934
          - name: Prediction source-token precision
            type: prediction_source_content_token_precision
            value: 0.863554
      - task:
          type: summarization
          name: Summarization
        dataset:
          name: Recent GDELT news (gdelt_lv)
          type: gdelt_lv
        metrics:
          - name: ROUGE-L F1
            type: rougeL
            value: 0.201971
          - name: Reference content-token coverage
            type: reference_content_token_coverage
            value: 0.217
          - name: Prediction source-token precision
            type: prediction_source_content_token_precision
            value: 0.813343
      - task:
          type: summarization
          name: Summarization
        dataset:
          name: Held-out supported news (news_supported_lv)
          type: news_supported_lv
        metrics:
          - name: ROUGE-L F1
            type: rougeL
            value: 0.28415
          - name: Reference content-token coverage
            type: reference_content_token_coverage
            value: 0.283575
          - name: Prediction source-token precision
            type: prediction_source_content_token_precision
            value: 0.816354

Latvian summariser

Rahvusarhiiv/lv_summariser is an experimental Latvian summarization model fine-tuned from facebook/mbart-large-50-many-to-many-mmt.

The default branch contains revision 2, published in July 2026. The previously released checkpoint is preserved under the v1 tag. Revision 2 improves supported-news and recent-news summarization, source grounding, and removal of learned EU legal boilerplate, while producing shorter encyclopedic summaries.

This is a research release. It is not claimed to be state of the art; evaluate it on the target domain before production use.

Intended Use

• Summarization of Latvian news, legal/administrative, encyclopedic, or academic-style text.
• Batch/offline summarization, evaluation, and retrieval-assisted workflows where outputs can be checked.
• Not intended as a sole source of factual truth, legal advice, medical advice, or archival description without human review.

How to Use

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

repo_id = "Rahvusarhiiv/lv_summariser"
language_code = "lv_LV"
text = "Latvijas pētnieki paziņoja par jaunu projektu, kura mērķis ir uzlabot digitālo arhīvu pieejamību. Projekts paredz sadarboties ar bibliotēkām, universitātēm un valsts institūcijām."

tokenizer = AutoTokenizer.from_pretrained(repo_id, src_lang=language_code)
model = AutoModelForSeq2SeqLM.from_pretrained(repo_id)

inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=768)
language_token_id = tokenizer.convert_tokens_to_ids(language_code)
summary_ids = model.generate(
    **inputs,
    forced_bos_token_id=language_token_id,
    max_new_tokens=512,
    num_beams=4,
    no_repeat_ngram_size=3,
)
print(tokenizer.batch_decode(summary_ids, skip_special_tokens=True)[0])

convert_tokens_to_ids is used deliberately. Current Transformers releases may load the tokenizer as TokenizersBackend, which does not expose the older tokenizer-specific lang_code_to_id attribute.

Recommended Generation Settings

• max_input_tokens: 768
• max_new_tokens: 512
• num_beams: 4
• length_penalty: 1.0
• no_repeat_ngram_size: 3

These are the settings used for the revision-2 comparison evaluation.

Training Data

Revision 2 was trained on 30,000 balanced examples after protected-benchmark exclusion:

family	training rows	share
Latvian news	27,180	90.6%
EU legal	2,412	8.0%
Academic/thesis	408	1.4%

The sources include:

• 27,143 accepted articles from the locally collected LSM archive and 37 additional LSM/Delfi RSS examples.
• 2,412 cleaned dennlinger/eur-lex-sum examples. FMISummarization/FMI_Summarization was used to recover supporting source acts and select evidence; it was not added as duplicate sampling mass.
• 408 Latvian academic OAI abstract/full-text pairs collected from Latvian university journal repositories.

Legal targets underwent more aggressive summary-section extraction and boilerplate filtering. Long inputs use source-span evidence selection. The build excluded protected Wiki4, GDELT, and supported-news benchmark overlap using IDs, normalized URLs, exact source/summary/pair hashes, and shared token shingles.

Training metadata:

• Base model: facebook/mbart-large-50-many-to-many-mmt
• Training rows: 30,000
• Validation rows: 708
• Max source tokens: 768
• Max target tokens: 384
• Seed: 20260713

Evaluation

Automatic evaluation used 100 examples per benchmark with the same decoding settings for revision 1 and revision 2.

benchmark	ROUGE-L	ROUGE-1	ROUGE-2	ref coverage	src precision	words	repeat 4g	legalese	max hit
Strict Wiki4	0.313	0.347	0.251	0.291	0.864	28.7	0.000	0.000	0.000
Recent GDELT news	0.202	0.240	0.104	0.217	0.813	21.0	0.000	0.000	0.000
Held-out supported news	0.284	0.331	0.176	0.284	0.816	24.4	0.000	0.000	0.000

Revision 1 Comparison

benchmark	v1 ROUGE-L	v2 ROUGE-L	v1 src precision	v2 src precision
Strict Wiki4	0.387	0.313	0.787	0.864
Recent GDELT news	0.160	0.202	0.798	0.813
Held-out supported news	0.154	0.284	0.779	0.816

A local gemma4:31b Ollama judge evaluated five paired random examples per benchmark. This small comparative sample is not an absolute quality estimate.

benchmark	v1 factuality	v2 factuality	v1 coverage	v2 coverage	v1 overall	v2 overall
Strict Wiki4	3.20	3.80	2.20	2.20	2.60	2.60
Recent GDELT news	3.00	2.80	1.80	2.80	2.00	2.40
Held-out supported news	2.20	3.20	1.80	2.40	1.60	2.60

Revision 2 was selected because it improved both automatic and judge results on the two news benchmarks, increased source precision in all three domains, and removed detected legalese artifacts. Wiki4 ROUGE and reference coverage declined because the new model produces shorter encyclopedic summaries; this is a known tradeoff rather than a universal improvement.

Limitations and Failure Modes

• The model can hallucinate or over-generalize details, especially when the input is long or under-specified.
• Named entities, numbers, dates, and legal references should be verified against the source.
• Revision 2 can under-summarize broad encyclopedic inputs and omit secondary points.
• Summary length control was present in parts of training, but this release does not expose a guaranteed length-control interface.
• Title-like opening phrases can still occur; the Wiki4 duplicate-title-prefix flag was 0.09 in this evaluation.
• The training data mixes sources with different writing styles and upstream terms; users are responsible for checking suitability for their use case.

Bias, Risks, and Ethical Considerations

The model reflects the distributions and editorial choices of its source corpora. Its revision-2 mix is heavily weighted toward Latvian public-broadcaster news, so its style and topic distribution are not representative of all Latvian writing. It may reproduce biases present in news, legal, and academic sources. Human review is recommended for public-facing summaries and high-impact decisions.

CO2 Emissions

Training emissions were not instrumented with CodeCarbon or an equivalent tracker. No numeric CO2 estimate is reported. Training was local fine-tuning of an mBART-50 checkpoint rather than pretraining from scratch.

Model Tree

• Base model: facebook/mbart-large-50-many-to-many-mmt
• Relationship: fine-tuned sequence-to-sequence summarization checkpoint.
• Previous release: v1

License

The model card uses license: other because the training mix contains multiple upstream data sources with different terms. Treat this as a research-use release and verify upstream data/license constraints before redistribution or commercial use.