My Hugging Face models and datasets, by theme (October 2026)

Every so often I look at my Hugging Face profile and realise it has quietly turned into an archive of whatever I was curious about that month. There are Whisper fine-tunes trained on my own voice, a few thousand photographs of Jerusalem, a library of system prompts, and (for reasons that made sense at the time) two image models of stuffed animals.

So here is a point-in-time snapshot, taken on 15 October 2026: 16 models and 72 datasets, all public. Rather than splitting them into a models pile and a datasets pile, I've grouped them by theme, because that is how they were actually made. A dataset usually exists because I wanted to train or test a model on something, and the two belong next to each other.

The live list is always at huggingface.co/danielrosehill. Expect this post to drift out of date fairly quickly (that's rather the point of dating it).

Hebrew-English code-switching

English speakers in Israel talk in "Hebrish": English sentences with Hebrew words dropped in, from makolet to Bituach Leumi. Speech models stumble on this constantly, so it is the theme I keep coming back to. The token classifier has its own write-up: Teaching a podcast to say makolet.

Models

• Whisper-Hebrish: Whisper fine-tuned on English sentences containing Hebrew words, with a demo Space

• hebrew-latin-token-classifier: XLM-RoBERTa token classifier that finds Hebrew words written in Latin letters inside English text

Datasets

• English-Hebrew-Mixed-Sentences: the original Hebrish set: English sentences with Hebrew interspersed, with audio

• English-Hebrew-Codeswitching-Sentences: a newer recorded set of English sentences with everyday Hebrew words and Israeli names, read aloud

• hebrew-latin-code-switching: text corpus with token-level BIO spans marking the Hebrew words; the training data for the classifier above

Speech recognition tuned to one voice

A long-running experiment in whether fine-tuning Whisper on my own voice (and on the technical vocabulary I actually dictate) beats stock models and commercial APIs. The answer is "sometimes", and the eval datasets record which times.

Models

• daniel_whisper_finetune_large_v3_turbo_v2: Whisper Large v3 Turbo fine-tuned on my voice

• daniel_whisper_finetune_medium_v2: the same, on Whisper Medium

• daniel_whisper_finetune_small_v2: the same, on Whisper Small

• daniel_whisper_finetune_base_v2: the same, on Whisper Base

• daniel_whisper_finetune_tiny_v2: the same, on Whisper Tiny

• daniel_whisper_acft_small_v2: ACFT variant of the Small fine-tune

• daniel_whisper_acft_base_v2: ACFT variant of the Base fine-tune

• daniel_whisper_acft_tiny_v2: ACFT variant of the Tiny fine-tune

Datasets

• Tech-Sentences-For-ASR-Training: TechVoice: read-aloud sentences dense with technical terms, for ASR fine-tuning

• Tech-Sentences: a cleaner re-recording of dictation-style sentences full of DevOps, cloud and AI vocabulary

• Read-Aloud-Voice-Sentences: general read-aloud sentences from one speaker, recorded on a phone

• Whisper-Fine-Tune-One-Shot-Eval: back-of-the-envelope comparison of local fine-tunes against commercial ASR APIs

• STT-Fine-Tune-Eval-101225: evaluation results for the fine-tunes against stock Whisper, December 2025

Evaluating speech-to-text and audio models

Before tuning anything you need to know how badly the off-the-shelf options do. These are mostly real voice notes, recorded in kitchens, streets and cars rather than studios.

Datasets

• STT-Voice-Notes-Evals: comparative evaluation of STT services on voice-note transcription

• Voice-Note-Audio: real voice recordings with transcripts and annotations

• Noisy-Voice-Notes: a year of everyday voice notes recorded in noisy environments

• ASR-WPM-And-Background-Noise-Eval: how speaking speed and background noise affect Whisper accuracy

• Small-STT-Eval-Audio-Dataset: 92 clips with ground truth, covering technical vocabulary and code-switching

• Podcast-ASR-Evaluation: ASR evaluation on podcast audio

• Audio-Understanding-Test-Set: 137 test prompts across 22 categories for multimodal audio-understanding models

• Audio-Understanding-Bitrate-Eval-0426: how MP3 bitrate affects transcription across every audio-input LLM on OpenRouter

• Sample-Voice-Context-Data: synthetic narrated career history, for testing voice-based context gathering

From dictation to usable text

Transcription is only half the job. The other half is turning a rambling voice note into an email, a prompt, or a task, and routing it to the right place.

Models

• Voice-Note-Intent-Classifier: planned multi-label classifier for routing voice notes by intent (no weights yet; the card documents the plan)

Datasets

• Transcription-Cleanup-Trainer: paired raw and cleaned transcripts for training a cleanup model

• Prompt-Separation: voice-typed prompts decomposed into asks, context chunks and notes

• Text-Transformation-Prompts-300525: short system prompts for reformatting dictated text into different styles

• Speech-To-Text-System-Prompts-2: system prompts for refining STT output with an LLM

• ai-generated-podcast-episodes: audio from AI-generated podcast episodes

Jerusalem and Israel, photographed

I take a lot of street photographs, and at some point it seemed a waste not to make them useful as training data. The two LoRAs generate images in the style of the city photos; the poster detector came out of tracking a single postering campaign around central Jerusalem.

Models

• Jerusalem-Images: Flux LoRA trained on West Jerusalem streetscapes from summer 2025

• Tel-Aviv-Street-Style: Flux LoRA for Tel Aviv street scenes

• jerusalem-poster-detector: YOLO11 fine-tune that spots specific poster designs in street photos

Datasets

• Jerusalem-Streetscapes: Jerusalem's changing urban fabric, by day and by night

• Jerusalem-High-Rise-Development: 56 photos of high-rise construction around the city

• Tel-Aviv-Pics: 53 photographs of Tel Aviv, originally gathered as reference for 3D and game work

• Israel-Photos: 369 photos from across Israel with LLM-generated captions and locations

• jerusalem-poster-detection: geotagged street photos with bounding boxes around campaign posters

Hebrew text in images

Two sides of the same question: can vision models read Hebrew in the wild, and can image models write it?

Datasets

• Hebrew-Language-Signage: photographs of Hebrew signage around Israel

• Hebrew-Image-Eval-111225: 12 text-to-image models tested on rendering Hebrew text correctly

Israeli civic and open data

Public data that exists but is awkward to get at, mostly because it is Hebrew-only or buried in a portal. Some of this fed earlier posts, including the Jerusalem Shabbat air quality findings, the open data changelog and JLM Shelters.

Datasets

• Israel-Alerting-Zones: Home Front Command alerting zones, as of May 2025 (Hebrew)

• Jerusalem-Emergency-Shelters-0925: public shelter locations in Jerusalem, September 2025

• Israel-Open-Data-Catalogue: bilingual snapshot of the entire data.gov.il catalogue

• Jerusalem-Air-Quality-Shabbat: a year of 5-minute readings from 12 stations, set up for weekday vs Shabbat analysis

• Ivory-Catalog-020226: computer hardware prices from an Israeli retailer, February 2026

Geopolitics

Structured records of things I was tracking or simulating.

Datasets

• Iran-Israel-War-2026: OSINT dataset of Iranian missile and drone attack waves across the "True Promise" operations

• Geopol-Forecaster-Predictions: predictions extracted from multi-agent geopolitical wargaming runs

Character LoRAs (yes, plushies)

Small, silly, and genuinely the best way I found to learn how LoRA training behaves: a character with a fixed look is much easier to judge than a style.

Models

• Dinosaur-Sloth: Flux LoRA of a sloth plushie in a dinosaur costume

• Herman-Poppleberry: Flux LoRA of a handmade donkey plushie from Storrs, Connecticut

Datasets

• Corn-The-Sloth: keyframe photos of Cornelius, a plush sloth

• Corn-Training-Set: a second training set of the same sloth

Testing vision and image models

Small, single-purpose sets for poking at specific failure modes.

Datasets

• Whiteboards: handwritten whiteboards for OCR accuracy and VLM "pseudotext" hallucination

• NVR-Entity-Recognition-Experiment: entity recognition training data for a home NVR camera setup

• accidental-and-low-quality-photos: accidental and blurry camera-roll shots, for training a filter

• Text-To-Image-Test-Prompts: evaluation prompts for text-to-image models

System prompts and agent configurations

The oldest corner of the profile. These started as my personal prompt library and were split out by subject; the main library was updated by automation for a while, and the August 2025 export is a frozen copy of 944 prompts.

Datasets

• System-Prompt-Library: the main system prompt library

• System-Prompt-Library-030825: point-in-time export of the library, August 2025

• Writing-System-Prompts: prompts for writing assistants

• Career-Related-System-Prompts: career guidance and job search prompts

• Geopolitical-System-Prompts: prompts for structured geopolitical analysis

• Data-Utils-System-Prompts: prompts for data-handling assistants

• Email-Management-System-Prompts: email drafting and triage prompts

• Prompt-Eng-System-Prompts: prompts for prompt engineering

• General-Purpose-System-Prompts: a few general defaults for when a chat UI has none

• Shakespearean-Text-Transformation-Prompts: possibly the internet's largest collection of prompts for rewriting text as Shakespeare

• Code-Gen-Agents-0925: subagent configurations for code generation, September 2025

• Narcissistic-Abuse-Support-Configs: AI agent configurations for people dealing with abusive relationships

• prompt-or-not: labelled examples for telling prompts apart from other text

Context data for AI tools

Experiments in giving models a durable sense of who I am, published as plain repositories an agent can read.

Datasets

• Software-Wish-List-Context-Data: my general software preferences, for stack evaluations

• Career-Data-Context-Repo: career context for agents and LLMs

• Blog-Posts: older blog posts I used to fine-tune a writing-style generator

• github-repos-100825: snapshot of my GitHub repositories, August 2025

LLM experiments and market snapshots

One-off experiments, plus a couple of catalogues captured on a particular day.

Datasets

• Single-Prompt-Book: can a model write a whole book from one prompt?

• Long-Prompt-Experiment: does prompt structure matter more than prompt length?

• ChatGPT-AI-Vs-API: comparing outputs from the ChatGPT app and the API

• Open-Router-API-Pricing-Analysis: pricing and parameters for every model on OpenRouter, November 2025

• multimodal-ai-taxonomy: a taxonomy of multimodal model capabilities by input and output modality

• Zapier-Integrations-260825: the Zapier integrations catalogue, August 2025

Impact and sustainability data

From my work around impact measurement: reshaped versions of public datasets that are hard to analyse in their original form.

Datasets

• Global-Value-Factor-Database-Refactor-V2: refactor of the IFVI Global Value Factor Database, V2

• ifvi_valuefactors_deriv: the earlier version, now deprecated in favour of V2

• GHG-Emissions-Data: a pipeline combining greenhouse gas emissions sources

• pay-for-outcomes-instruments: social impact bonds and outcomes contracts, from the Oxford GO Lab data

• ISO-3166-4217-Consolidated: one lookup table joining countries to currencies

Physical things

The newest theme, from trying to organise a home (and a storage unit) with something more rigorous than labelled boxes.

Datasets

• storage-container-dimensions: planning-grade dimensions for Euroboxes and other industrial containers

• homegoods-taxonomy: a hierarchical classification of household goods, with handling attributes

What's not here

Hugging Face Spaces (about 50 of them, mostly demos and small tools) are left out, as are anything private. If one of these is useful to you, or you'd like a licence clarified, the easiest route is a discussion on the repo itself.