hypaai/Hypa-Keyboard-v1

Hypa Keyboard v1 is a 409,598-example instruction dataset for training on-device smart keyboard models — next-word prediction, word completion, autocorrect, and grammatical error correction — across 27 languages, weighted heavily toward African languages that no mainstream keyboard supports.

Every example is a three-turn chat (system → user → assistant), so the dataset can be fed directly to any chat-template SFT pipeline. Errors in the input text are synthetically injected from a controlled corruption vocabulary, and each example carries metadata describing exactly which corruptions were applied, how many, and how severe the result is.

Built by Hypa AI.

---

Why this dataset

Mobile keyboards for African languages are effectively unserved. Typing Igbo, Yorùbá, Efik, Tiv, Igede or Eggon on a mainstream keyboard means fighting autocorrect rather than being helped by it — diacritics get stripped, valid words get "corrected" into English, and there is no next-word prediction at all. Of the 27 languages here, only a handful have any keyboard support worth the name, and several have no published NLP resources at all.

Hypa Keyboard v1 targets the four tasks a keyboard actually needs to perform, in the languages that mainstream keyboards ignore. English, French, Spanish, Portuguese and Arabic are included both because they are genuinely code-switched into across these regions and because they give a model anchoring signal in higher-resource text.

The dataset explicitly models tone-mark damage (global_strip_tone, swap_tone_word) as a first-class error type, because losing diacritics is the single most common failure mode when typing tonal orthographies on a standard mobile keyboard.

has five classes, each paired with a fixed system prompt:

task	System prompt	What the model must do
predict	You are Hypa Keyboard. Predict the next word.	Given text ending at a word boundary, emit the next word
complete	You are Hypa Keyboard. Complete the current word.	Given text ending mid-word, finish the word being typed
correct	You are Hypa Keyboard. Correct the last word.	Fix only the final (just-typed) token
correct_block	You are Hypa Keyboard. Correct the text block.	Return a clean version of the whole span
grammar	You are Hypa Keyboard. Correct grammar, missing words, spelling, and sentence errors.	Full grammatical error correction over the span

predict and complete examples are always drawn from clean text (n_edits = 0, difficulty = "clean"). The three correction tasks carry injected noise.

---

Languages

27 languages, grouped by region:

Nigeria — Annang, Ebira, Efik, Eggon, Fulani, Hausa, Ibibio, Idoma, Igala, Igbo, Igede, Nigerian Pidgin, Nupe, Tiv, Urhobo, Yorùbá

Wider Africa — Chichewa, Fon (Fongbe), Kirundi, Somali, Swahili, Twi

Global — Arabic, English, French, Portuguese, Spanish

The majority of these are low-resource languages with little or no existing keyboard-assistance support. Several — Eggon, Igede, Ebira, Annang, Nupe — have almost no prior NLP resources of any kind.

Language	Code	Language	Code	Language	Code
Annang	anw	Hausa	ha	Portuguese	pt
Arabic	ar	Ibibio	ibb	Somali	so
Chichewa	ny	Idoma	idu	Spanish	es
Ebira	igb	Igala	igl	Swahili	sw
Efik	efi	Igbo	ig	Tiv	tiv
Eggon	ego	Igede	ige	Twi	tw
English	en	Kirundi	rn	Urhobo	urh
Fon	fon	Nigerian Pidgin	pcm	Yorùbá	yo
French	fr	Nupe	nup
Fulani	ff

Language identity is not stored as a column, and several of these languages are close enough in vocabulary and orthography (Efik / Ibibio / Annang, for instance) that automatic identification is unreliable between them. If you need per-language filtering, the practical route is to recover it from the source corpus at generation time rather than post-hoc LID. Adding a lang column is the single highest-value improvement for v3.

---

Dataset structure

Size

Rows	409,598
Splits	train only
Format	Parquet
On-disk size	27.6 MB

Fields

Column	Type	Description
messages	list[dict] (length 3)	The conversation. Each message has role, content, thinking, tool_calls, tool_responses. Roles are always system, user, assistant in that order. thinking is "" and the tool fields are null throughout — they exist for schema compatibility with tool-calling pipelines.
task	string	One of predict, complete, correct, correct_block, grammar.
n_edits	int32	Number of corruption operations applied to the user text. Range 0–18.
edit_types	list[string]	Which corruption operations were applied (deduplicated, sorted). Empty for clean examples. Up to 11 distinct types per example.
edit_ratio	float32	Severity of corruption, 0.0–1.0 — roughly the fraction of the span affected. Useful for curriculum scheduling or filtering out unrecoverably mangled examples.
difficulty	string	clean, easy, medium, or hard, derived from edit_ratio and span length.
span_words	int32	Word count of the input span. Range 0–96.
has_tone	bool	Whether the target text contains tone marks / diacritics. Lets you evaluate tonal and non-tonal orthographies separately.

Corruption vocabulary (edit_types)

Word-level word_typo · duplicate_word · delete_word · swap_words · merge_words · split_word

Casing decap_word · decap_initial · capslock_word · global_lowercase

Punctuation & whitespace punct_drop · punct_insert · punct_swap · punct_double · double_space

Tone / diacritics global_strip_tone · swap_tone_word

Example

{
  "messages": [
    {"role": "system",    "content": "You are Hypa Keyboard. Correct the text block.", "thinking": "", "tool_calls": null, "tool_responses": null},
    {"role": "user",      "content": "Omi Omi kp anya'ami mail ekubo uche r'abo ohigbeli mi.", "thinking": "", "tool_calls": null, "tool_responses": null},
    {"role": "assistant", "content": "<corrected text>", "thinking": "", "tool_calls": null, "tool_responses": null}
  ],
  "task": "correct_block",
  "n_edits": 3,
  "edit_types": ["duplicate_word", "global_strip_tone", "swap_words"],
  "edit_ratio": 0.222,
  "difficulty": "medium",
  "span_words": 11,
  "has_tone": true
}

---

Usage

from datasets import load_dataset

ds = load_dataset("hypaai/Hypa-Keyboard-v1", split="train")

# Task mix
print(ds.to_pandas()["task"].value_counts())

# Train only the correction head
corrections = ds.filter(lambda x: x["task"] in {"correct", "correct_block", "grammar"})

# Curriculum: start easy, ramp up
easy = ds.filter(lambda x: x["edit_ratio"] <= 0.15)

# Tonal-orthography evaluation slice
tonal = ds.filter(lambda x: x["has_tone"])

# Drop examples corrupted past the point of recovery
usable = ds.filter(lambda x: x["edit_ratio"] < 0.6)

SFT with TRL

from trl import SFTTrainer

trainer = SFTTrainer(
    model="<your-base-model>",
    train_dataset=ds,          # `messages` is already in chat format
    args=...,
)
trainer.train()

---

Dataset creation

Source data

Target (clean) text is drawn from Hypa AI's multilingual corpora, which include machine-translation outputs, Wikipedia-derived content, dictionary and definition entries, and instruction-formatted text. Some spans are mid-document slices, so a fraction of examples retain artefacts of their source formatting — Markdown markers (**, ###), fragments of translation instructions (Return ONLY THE EXACT TRANSLATION), and JSON punctuation appear in the input text.

This is partly intentional: a real keyboard sees messy input. But it also means the dataset is not a clean monolingual corpus, and a model trained on it will occasionally see prompt-like text in the typing stream.

Annotation process

Corruptions are programmatically injected — no human annotation. For each noised example the generator samples a set of operations from the corruption vocabulary, applies them to clean target text, and records the operation set, count, and resulting severity. The assistant turn always holds the original clean text (or the correct continuation, for the prediction tasks).

Personal and sensitive information

Source text is drawn from public and machine-generated corpora and includes named public figures, organisations and places, as encountered in that source material. No private user keystroke data was collected or used.

---

Limitations and bias

• No language column. Per-language training, evaluation and mixture control all require an extra LID step.
• Synthetic noise ≠ real typing errors. The corruption vocabulary does not model keyboard-layout adjacency (fat-finger errors), swipe-typing failure modes, or code-switching mid-sentence — all of which dominate real African-language mobile input.
• No dedicated validation or test split. Hold out your own, and split by source document rather than randomly if you want an honest generalisation estimate.
• Uneven language coverage. Across 27 languages the mix is certainly not balanced — high-resource languages and the better-documented Nigerian languages will dominate, while Eggon, Igede, Ebira and Nupe are likely thin. Without a lang column you cannot verify or correct this, so do not assume uniform performance across the list.
• Close-relative confusion. Efik, Ibibio and Annang share substantial vocabulary and orthography, as do several of the Benue-Congo languages here. A model may transfer corrections across them in ways that are wrong for a given speaker.
• Code-switching. Nigerian Pidgin and English blend continuously in real typing; the dataset treats them as separate sources rather than modelling the mixture explicitly.
• Source-formatting leakage. As noted above, some spans carry Markdown, prompt fragments, or JSON syntax.
• Extreme corruption. At high edit_ratio (≥ 0.5 on short spans) the input can be unrecoverable even for a human. Filter these out unless you are deliberately training robustness.
• Machine-translated source text carries its own errors, which the model will learn as correct targets in the affected languages.

---

Intended uses

Suitable for: fine-tuning small on-device models for next-word prediction, autocomplete, autocorrect and GEC; benchmarking keyboard-assistance quality across African languages; research on diacritic restoration.

Not suitable for: general-purpose language modelling, translation training, or any use treating the assistant turns as authoritative reference text for the languages covered.

---

Citation

@misc{hypaai2025keyboardv1,
  title        = {Hypa Keyboard v1: A 27-Language Instruction Dataset for African-Language Smart Keyboards},
  author       = {Hypa AI},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/hypaai/Hypa-Keyboard-v1}}
}

License

Released under the Apache License 2.0. Free to use, modify, and redistribute for both research and commercial purposes. Component sources include human-translated content from the Hypa Intelligence and AfroVoices teams, augmented parallel data, and synthetic instructions; all components were released by their originating teams under licenses compatible with Apache 2.0 redistribution.

Contact

Hypa Intelligence • Website • Hugging Face • GitHub • Blog