Duplicate from Adedoyin00/yoruba-pronoun-etiquette

Co-authored-by: Adedoyin Christianah Bamigbala <Adedoyin00@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,61 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.avro filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.lz4 filter=lfs diff=lfs merge=lfs -text
+*.mds filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+# Audio files - uncompressed
+*.pcm filter=lfs diff=lfs merge=lfs -text
+*.sam filter=lfs diff=lfs merge=lfs -text
+*.raw filter=lfs diff=lfs merge=lfs -text
+# Audio files - compressed
+*.aac filter=lfs diff=lfs merge=lfs -text
+*.flac filter=lfs diff=lfs merge=lfs -text
+*.mp3 filter=lfs diff=lfs merge=lfs -text
+*.ogg filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
+# Image files - uncompressed
+*.bmp filter=lfs diff=lfs merge=lfs -text
+*.gif filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.tiff filter=lfs diff=lfs merge=lfs -text
+# Image files - compressed
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.jpeg filter=lfs diff=lfs merge=lfs -text
+*.webp filter=lfs diff=lfs merge=lfs -text
+# Video files - compressed
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.webm filter=lfs diff=lfs merge=lfs -text
+yoruba_pronoun_etiquette.csv filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,301 @@
+---
+language:
+- yo
+license: other
+pretty_name: Yoruba Honorific Pronoun Etiquette Benchmark
+size_categories:
+- 1K\<n\<10K
+tags:
+- yoruba
+- linguistics
+- pragmatics
+- politeness
+- honorifics
+- pronouns
+- low-resource-language
+- african-languages
+- llm-evaluation
+- cultural-alignment
+---
+
+# Yoruba Honorific Pronoun Etiquette Benchmark
+
+## Overview
+
+This dataset is a Yoruba-language benchmark focused on **honorific
+pronouns, politeness markers, and socially appropriate pronoun
+selection**.
+
+The project is motivated by the observation that politeness is
+culturally and socially grounded. In Yoruba, choices among second-person
+forms can encode age, social status, familiarity, deference, and
+interpersonal relationships. An LLM may produce a grammatically possible
+Yoruba form while still failing to preserve the culturally appropriate
+level of respect.
+
+The aim of this project is therefore not only to test Yoruba language
+generation, but to evaluate whether a language model can maintain
+**socially appropriate Yoruba pronoun choices** when speaker-addressee
+relationships and discourse contexts change.
+
+## Dataset contents
+
+The repository contains both the raw/source dataset and the
+AdaptLab-adapted dataset, together with benchmark and error-analysis
+outputs.
+
+### Main datasets
+
+-   `yoruba_pronoun_etiquette.csv` --- raw/source dataset.
+-   `Yoruba_5000_AdaptLab_READY.csv` --- 5,000-row AdaptLab-adapted
+    dataset.
+
+### Benchmark outputs
+
+-   `yoruba_pronoun_benchmark_clean.csv` --- cleaned benchmark data.
+-   `yoruba_pronoun_benchmark_v4.csv` --- V4 benchmark results.
+-   `yoruba_pronoun_results_v4.csv` --- pronoun-level results.
+-   `yoruba_relationship_results_v4.csv` --- results grouped by social
+    relationship.
+-   `yoruba_pronoun_error_matrix_v4.csv` --- pronoun error matrix.
+
+## Raw dataset
+
+The raw dataset contains controlled Yoruba examples designed around
+pronoun etiquette and social relationships. The examples encode
+contextual information such as:
+
+-   speaker/addressee relationship;
+-   relative age or social position;
+-   familiarity;
+-   communicative situation;
+-   candidate pronoun forms;
+-   target politeness/honorific interpretation.
+
+The raw dataset is a controlled dataset rather than a collection of
+naturally occurring conversations. This distinction is important when
+interpreting the results.
+
+## Adapted dataset
+
+The adapted dataset was produced using **AdaptLab**. The adaptation
+process reformulated the original examples into model-oriented
+prompt/completion examples while retaining the central linguistic and
+pragmatic task.
+
+The adapted data includes Yoruba contexts in which the model is required
+to select or produce an appropriate form based on the relationship
+between interlocutors.
+
+Examples include contrasts involving relationships such as:
+
+-   peer → peer;
+-   younger → elder;
+-   younger → older visitor;
+-   younger → leader;
+-   student → teacher;
+-   student → professor;
+-   adult → child;
+-   younger → family head;
+-   younger → Babaláwo;
+-   close colleagues;
+-   friends;
+-   siblings/close family.
+
+## Annotation scheme
+
+The benchmark treats pronoun selection as a pragmatic
+classification/evaluation problem rather than purely a surface-form
+matching task.
+
+The relevant annotation dimensions include:
+
+1.  **Social relationship** --- who is speaking to whom.
+2.  **Discourse context** --- the situation in which the utterance
+    occurs.
+3.  **Candidate forms** --- the pronoun alternatives presented to the
+    model.
+4.  **Gold pronoun** --- the expected socially appropriate form.
+5.  **Adapted pronoun** --- the form produced/selected after adaptation.
+6.  **Politeness/honorific function** --- whether the form reflects
+    deference, solidarity, or a non-honorific usage.
+
+The project distinguishes forms including `ìwọ`, `ẹ`, `ẹ̀yin`, `yín`,
+`o`, `ọ`, and `rẹ` where they occur in the benchmark.
+
+## Cultural and linguistic motivation
+
+Politeness is treated in this project as a culturally situated component
+of language use. Yoruba honorific usage cannot be reduced to a simple
+grammatical rule because social variables such as age, status,
+familiarity, and relationship influence appropriate forms.
+
+The benchmark therefore deliberately places pronoun choices inside
+social contexts. This is intended to test whether an LLM preserves
+culturally meaningful distinctions rather than simply generating a
+locally grammatical Yoruba expression.
+
+The project follows the linguistic literature supplied with the
+research, including work discussing Yoruba honorific pronouns, age,
+social status, familiarity, and the relationship between politeness and
+cultural norms.
+
+## AdaptLab and model adaptation
+
+AdaptLab was used as the dataset adaptation tool.
+
+The adapted dataset was subsequently used in model
+adaptation/fine-tuning experiments. The AutoScientist workflow was used
+for the training experiment.
+
+The reported training configuration included:
+
+-   base model: `meta-llama/Llama-4-Scout-17B-16E-Instruct`;
+-   training method: supervised fine-tuning (SFT);
+-   parameter-efficient training: LoRA;
+-   LoRA rank: 64;
+-   LoRA alpha: 128;
+-   learning rate: 0.0001;
+-   epochs: 4;
+-   batch size: max;
+-   learning-rate scheduler: cosine;
+-   warmup ratio: 0.03;
+-   weight decay: 0.02;
+-   dropout: 0;
+-   training on inputs: false.
+
+The trained model was named:
+
+`adaption_yoruba_pronoun_etiquette`
+
+## Benchmark methodology
+
+The final V4 benchmark independently compares the adapted model output
+with the expected/gold pronoun.
+
+From the V4 benchmark run:
+
+-   total source rows inspected: 27,780;
+-   relationships extracted: 2,096;
+-   gold pronouns identified: 2,093;
+-   adapted pronouns identified: 2,066;
+-   valid benchmark cases: 2,064;
+-   correct predictions: 1,705;
+-   incorrect predictions: 359;
+-   exact pronoun accuracy: **82.61%**.
+
+The benchmark also reports results by social relationship and by gold
+pronoun, together with an error matrix.
+
+## Key benchmark findings
+
+Overall exact pronoun accuracy was **82.61%**.
+
+Performance was strongest for several explicitly hierarchical
+relationships. For example:
+
+-   younger → elder: 98.94%;
+-   adult → child: 100%;
+-   younger → older visitor: 100%;
+-   younger → elder neighbour: 100%;
+-   younger → leader: 100%;
+-   younger → older man: 100%;
+-   younger → older woman: 100%;
+-   younger → family head: 100%;
+-   student → teacher: 100%;
+-   student → professor: 100%;
+-   younger → Babaláwo: 100%.
+
+The largest relationship category was peer → peer, with 1,767 valid
+cases and 80.08% accuracy. This is important because peer interactions
+create greater ambiguity between solidarity/informal forms and honorific
+forms.
+
+By gold pronoun, the V4 benchmark reported:
+
+-   `ìwọ`: 72.47%;
+-   `o`: 87.38%;
+-   `ọ`: 81.05%;
+-   `ẹ̀yin`: 97.73%;
+-   `yín`: 100%;
+-   `ẹ`: 97.53%;
+-   `rẹ`: 100%.
+
+The benchmark also identified recurring confusions among candidate
+forms, especially in peer-to-peer contexts.
+
+## Error analysis
+
+The principal error pattern is not simply random Yoruba generation. Many
+errors involve choosing a different candidate pronoun from the set
+supplied in the prompt.
+
+For example, the V4 error analysis contains cases where:
+
+-   gold = `ọ`, adapted = `ẹ`;
+-   gold = `ìwọ`, adapted = `yín`;
+-   gold = `o`, adapted = `ẹ`;
+-   gold = `ìwọ`, adapted = `ẹ̀yin`.
+
+The concentration of errors in peer → peer cases suggests that socially
+less hierarchical interactions can be more difficult to disambiguate
+than strongly marked age/status relationships.
+
+This supports the central research motivation: **culturally appropriate
+Yoruba politeness requires sensitivity to social context, not only
+lexical or grammatical competence.**
+
+## Intended use
+
+This dataset is intended for:
+
+-   evaluation of Yoruba-capable language models;
+-   research on cultural and pragmatic alignment;
+-   low-resource African-language NLP;
+-   evaluation of honorific and politeness-sensitive generation;
+-   research into socially grounded pronoun selection;
+-   benchmarking model adaptation methods.
+
+It should not be treated as a comprehensive representation of all Yoruba
+speakers, dialects, communities, or politeness practices.
+
+## Limitations
+
+1.  The raw benchmark data are controlled/constructed examples rather
+    than a corpus of naturally occurring conversations.
+2.  Yoruba politeness is context-sensitive, and the benchmark cannot
+    represent every possible social interaction.
+3.  Some pronoun forms are multifunctional and may require discourse
+    information beyond a short prompt.
+4.  Social categories such as age, status, familiarity, and relationship
+    may overlap in real interaction.
+5.  Benchmark accuracy should therefore be interpreted as performance on
+    the defined evaluation task, not as a complete measure of Yoruba
+    language competence.
+6.  Cultural and pragmatic judgments may vary across speakers and
+    communities.
+
+## Reproducibility
+
+The repository provides the raw and adapted datasets together with
+benchmark outputs so that researchers can inspect the transformation and
+reproduce the reported evaluation.
+
+The benchmark results were generated with the V4 evaluation pipeline and
+saved as CSV files in this repository.
+
+## Citation
+
+If you use this dataset, please cite the associated competition
+submission/research paper:
+
+> Bamigbala Christianah Adedoyin. *Yoruba Honorific Pronoun Etiquette:
+> Culturally Grounded Pronoun Adaptation and Benchmarking for LLMs*.
+> 2026.
+
+## Acknowledgement
+
+This project was developed as part of an African computational
+linguistics/LLM adaptation research activity focused on improving the
+representation of Yoruba pragmatic and cultural knowledge in language
+models.

yoruba_pronoun_error_matrix_v4.csv

@@ -0,0 +1,9 @@
+Gold,o,rẹ,yín,Ìwọ,ìwọ,Ẹ,ẹ,ẹ̀yin,ọ
+o,526,0,14,0,0,0,38,24,0
+rẹ,0,14,0,0,0,0,0,0,0
+yín,0,0,84,0,0,0,0,0,0
+ìwọ,0,0,35,1,437,1,84,45,0
+Ẹ̀yin,0,0,0,0,0,0,0,1,0
+ẹ,0,0,0,0,0,0,79,1,1
+ẹ̀yin,0,0,0,0,1,0,1,86,0
+ọ,0,0,27,0,0,0,51,34,479

yoruba_pronoun_etiquette.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:31ec935908e1031d36f4a616a1786a2c626e5740a06bd2bf03a1973a7b644c0b
+size 145858133

yoruba_pronoun_results_v4.csv

@@ -0,0 +1,9 @@
+gold_pronoun,samples,correct,accuracy
+ìwọ,603,437,72.4709784411277
+o,602,526,87.37541528239203
+ọ,591,479,81.04906937394247
+ẹ̀yin,88,86,97.72727272727273
+yín,84,84,100.0
+ẹ,81,79,97.53086419753086
+rẹ,14,14,100.0
+Ẹ̀yin,1,0,0.0

yoruba_relationship_results_v4.csv

@@ -0,0 +1,24 @@
+relationship,samples,correct,accuracy
+peer → peer,1767,1415,80.07923033389926
+younger → elder,189,187,98.94179894179894
+adult → child,8,8,100.0
+younger → older visitor,8,8,100.0
+younger → elder neighbour,8,8,100.0
+friends,7,6,85.71428571428571
+younger → leader,7,7,100.0
+siblings/close family,7,6,85.71428571428571
+peers,7,6,85.71428571428571
+younger → older man,7,7,100.0
+younger → older woman,6,6,100.0
+close colleagues,6,6,100.0
+younger → family head,6,6,100.0
+student → professor,5,5,100.0
+student → teacher,5,5,100.0
+younger → Babaláwo,5,5,100.0
+young adult → close friend,4,3,75.0
+adult → familiar child,3,3,100.0
+adult → younger relative,3,3,100.0
+close friends,2,2,100.0
+classmates,2,1,50.0
+siblings/close family — younger sibling,1,1,100.0
+siblings/close family — older sibling,1,1,100.0