For the REST-MEX 2026 edition, the evaluation follows a fully generative and indirect assessment framework. Unlike traditional classification tasks, participants do not train or evaluate their models directly on real test partitions.

Participants are provided with a restricted set of real tourist reviews exclusively for training or adapting their generative models. Using this material, each team must generate a synthetic dataset consisting of 3,000 original tourist reviews, each one explicitly labeled with a sentiment polarity value.

To evaluate participant submissions, the organizers will train a sentiment classification model using only the synthetic data provided by each team. No real reviews or external data sources are used during this training phase.

Once the classifier has been trained on the synthetic corpus, it is evaluated on a hidden test partition composed entirely of real tourist reviews. This test set is not accessible to participants and serves to measure how well the synthetic data supports generalization to real-world sentiment analysis.

The performance of the classifier on the hidden real test set determines the final score. Under this evaluation scheme, the effectiveness of each participant’s approach is measured by the impact of the generated synthetic data on downstream sentiment classification, rather than by the intrinsic quality of the generated texts alone

• C
  The set of sentiment polarity classes:
  C={1 (very negative),2 (negative),3 (neutral),4 (positive),5 (very positive)}C = \{1\text{ (very negative)}, 2\text{ (negative)}, 3\text{ (neutral)}, 4\text{ (positive)}, 5\text{ (very positive)}\}C={1 (very negative),2 (negative),3 (neutral),4 (positive),5 (very positive)}

• Fi(k)
  The F1-score obtained by system kkk for polarity class iii.
  This measures how well the classifier identifies that specific sentiment class.

• TCi
  The total number of real reviews belonging to class iii in the evaluation corpus.

• TC
  The total number of reviews across all polarity classes in the real corpus.

For REST-MEX 2026, participants do not submit classification runs. Instead, each team must submit a synthetic dataset generated by their proposed model.

• Submission window: until April 24, 2026 (23:59, UTC-06:00)

• Each participating team is allowed to submit one synthetic dataset, which must consist of exactly 3,000 original tourist reviews, each explicitly labeled with a sentiment polarity (from 1 to 5).

• Once the submission period is closed, no further modifications or additional submissions will be accepted. The organizers will then use each submitted dataset as the exclusive training data for the sentiment classifier used in the official evaluation.

• The final ranking will be determined solely by the performance obtained on the hidden real test set, using the official RP evaluation metric.

For REST-MEX 2026, participants do not submit classification outputs or prediction files. Instead, each team must submit one or more synthetic datasets generated by their proposed model.

Submission format

Submissions must be provided as a CSV file with the following strict requirements:

• The file must contain exactly two columns:

  • opinion — the synthetic tourist review text (in Spanish)

  • polarity — the sentiment label assigned by the system (integer value from 1 to 5)

• The CSV file must contain at most 3,001 rows:

  • 1 header row

  • Up to 3,000 synthetic reviews

• No missing values are allowed:

  • Every row must include both a valid opinion text and a polarity value.

  • Files containing empty cells or incomplete rows will be considered invalid.

• All reviews must be fully synthetic:

  • Generated texts must be original.

  • Copying, paraphrasing, or reusing real opinions from the training data is strictly prohibited.

Polarity values

The polarity column must contain one of the following integer values:

• 1 — Very negative

• 2 — Negative

• 3 — Neutral

• 4 — Positive

• 5 — Very positive
  
  

Values outside this range will invalidate the submission.

Example (CSV format)

opinion,                                                       polarity

"El lugar estaba descuidado y la atención fue muy deficiente.",1

"Aunque es bonito, la experiencia no cumplió con lo esperado.",2

"El sitio es correcto, sin nada particularmente destacable.",  3

"Nos gustó mucho el ambiente y la atención del personal.",     4

"Una experiencia excelente, sin duda regresaríamos.",          5

File naming and submission method

• File naming is flexible. We recommend using a clear identifier such as the team name, corresponding author, or submission version.

• Submissions must be sent via email to:📧 miguel.alvarez@cimat.mx

Important notes

• Multiple submissions per team are allowed.
  Each submission will be treated as an independent synthetic dataset.

• All submissions must comply with the format and content rules described above.

• Submissions that fail format validation or content requirements will be considered invalid.

• No modifications will be accepted after the official submission deadline.

This submission procedure ensures a transparent and reproducible evaluation of synthetic data generation strategies within REST-MEX 2026.