datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
tr-hi-mimi-encoded
TR↔HI Mimi-Encoded Parallel Speech
Pre-encoded parallel Turkish↔Hindi speech pairs for training speech-to-speech translation models. All audio has been tokenized through the Mimi neural audio codec (8 codebooks, 12.5 Hz, 24kHz) and stored as .pt files with word-level text alignments.
Dataset Summary
Source audio
~911 hours of synthetic parallel TR↔HI speech from tr-hi-parallel-speech-v2
TTS model
OmniVoice (voice design mode, 14 voice designs)… See the full description on the dataset page: https://huggingface.co/datasets/tiny-aya-translate/tr-hi-mimi-encoded.fleurs-tr-hi-mimi-encoded
fleurs-tr-hi-mimi-encoded
Mimi-encoded Turkish↔Hindi parallel speech pairs for TinyAya Stage 2
speech-to-speech translation training.
Contents
encoded/*.pt — 9212 Mimi-encoded audio pairs (kyutai/mimi, 8 codebooks, 12.5 Hz, 24 kHz).
Each file keys: pair_id, src_lang, tgt_lang, src_text, tgt_text, src_codes[8, T_src], tgt_codes[8, T_tgt].
encoded/*.alignments.json — 18424 Whisper word-level alignment sidecars
(.src.alignments.json / .tgt.alignments.json).… See the full description on the dataset page: https://huggingface.co/datasets/tiny-aya-translate/fleurs-tr-hi-mimi-encoded.tiny-aya-translate-hinglish-casual-stripped
Dataset Card for tiny-aya-translate-hinglish-casual-stripped
Dataset Summary
tiny-aya-translate-hinglish-casual-stripped is a lightweight, text-only derivative of the original tiny-aya-translate/hinglish-casual dataset.
The original dataset is designed for simultaneous translation and contains many columns including audio references, speaker metadata, and duration. It also includes paralinguistic tags (e.g., <sigh>, <laugh>, <chuckle>) embedded within the… See the full description on the dataset page: https://huggingface.co/datasets/bingbangboom/tiny-aya-translate-hinglish-casual-stripped.tiny-aya-base-blind-spots
Blind Spots of a Frontier Base Model: Evaluation Dataset
This dataset documents blind spots discovered in a frontier open-weight base model through 19 structured evaluation tests. It was assembled as part of an assignment on identifying model weaknesses using the HelloBench evaluation framework.
Model Tested
CohereLabs/tiny-aya-base
Architecture: Transformer with Sliding Window Attention (SWA) (window size 4096, with RoPE) on three layers + one global attention layer… See the full description on the dataset page: https://huggingface.co/datasets/Mawube/tiny-aya-base-blind-spots.tiny-aya-medical-concept-probes
Tiny Aya Cross-Lingual Medical Concept Probes
Dataset Description
20 medical concepts expressed as full sentences in 10 languages, designed for probing cross-lingual concept representations in multilingual LLMs. Each concept is a complete declarative sentence preserving the same semantic structure across all languages.
Purpose
These probe sentences serve as stimuli for mechanistic interpretability analysis -- specifically, extracting residual stream activations… See the full description on the dataset page: https://huggingface.co/datasets/s4um1l/tiny-aya-medical-concept-probes.toxigen_tiny-portuguesetiny-aya-base-blindspots
tiny-aya-base Blind Spot Dataset (10 probes)
This is a small, hand-curated evaluation set of 10 diverse prompts that surface
failure modes of the base model
CohereLabs/tiny-aya-base.
Every row contains the prompt prefix, the expected answer, and the model’s actual
output (greedy decoding).
The goal isn’t to be exhaustive—it’s to capture where the model is most brittle:
small arithmetic, ratios, formal negation, calendar logic, code errors, and
low‑resource language meaning.… See the full description on the dataset page: https://huggingface.co/datasets/KarmaIncarnate/tiny-aya-base-blindspots.tiny-aya-global-blindspots
tiny-aya-global-blindspots
Title & Overview
A curated set of failure cases for CohereLabs/tiny-aya-global, showcasing blind spots discovered while probing the ~3.35B-parameter base checkpoint released in February 2026. Each entry captures a prompt, the expected aligned behavior, and the model's actual output. The dataset illustrates common failure patterns observed when probing the ~3.35B multilingual base model without additional alignment or safety fine-tuning.… See the full description on the dataset page: https://huggingface.co/datasets/kaustubhg73/tiny-aya-global-blindspots.tiny-aya_failure
Dataset Summary
A compact diagnostic benchmark for evaluating failure modes in language models. It tests a model's ability to resist false premises, avoid fabricating entities, detect contradictions, follow strict constraints, and recognize category mismatches.
Each record includes: id, category, prompt, model_output, and expected_correct_output.
Model
Outputs were generated using CohereLabs/tiny-aya-global via the Hugging Face transformers library.
Generation settings:… See the full description on the dataset page: https://huggingface.co/datasets/akotet08/tiny-aya_failure.