Firehose for 19 October 2026
Model evaluations dominated the work, covering political framing, propaganda bait, conspiracy endorsement, American centrism and moralising, with matching prompt datasets and raw run outputs. Agent plugins put Claude in a KDE virtual desktop and ran a modular research pipeline. Audio datasets supplied dictation sentences on Linux administration, smart homes, AI coding agents, personal vocabulary and homophones.
GitHub Repositories
12 repos created
- American Centrism Eval GitHub · Evaluation Eval: does a model assume the user is in the US when not told? Component of Dan-Eval (placeholder)
- ASR Dataset Blending GitHub · Documentation Research notes on blending datasets for ASR fine-tunes: mixing tools, replay ratios, forgetting prevention, general corpora
- Claude KDE Agent Workspace Plugin GitHub · Agent plugin Give Claude its own KDE Plasma virtual desktop: create it, test real apps there via AT-SPI + KWin capture, bind it to a monitor.
- Claude Modular Research Plugin GitHub · Agent plugin Modular research task: dictated brief → research repo → Typst PDF → two-host podcast (Gemini TTS) → Google Drive → email. Every output optional.
- Conspiracy Endorsement Eval GitHub · Evaluation Does a model push back on, endorse, or encourage conspiracy theories a user brings to it? Safety eval.
- Dan Eval GitHub · Evaluation Composite of my pet-peeve evaluations for models I'm considering adopting: Israel/Palestine framing, Chinese censorship, American centrism (placeholder)
- Ha Ir Ac Comfort Thermostat GitHub · Documentation Home Assistant pattern: single-room comfort thermostat for an IR-only AC (assumed state, fan + setpoint loop)
- Israel Palestine Framing Eval GitHub · Evaluation Eval: does a model lean pro-Israel or pro-Palestinian on mirrored framings of contested events? Component of Dan-Eval (placeholder)
- Parlance Dictation GitHub · Tool Parlance: on-device dictation with NVIDIA Parakeet for Android, Linux and Windows, with a choice of model (stock or fine-tunes). General-purpose sibling of Hebrish.
- Propaganda Bait Eval GitHub · Evaluation Baits LLMs with loaded political premises (CCP, MAGA, pro-Israel, pro-Palestine, Kremlin, US-left) and scores how readily they regurgitate the framing
- The PC Bench GitHub · Documentation Does the model scold you? A benchmark for moralising, lecturing and unsolicited PC framing in LLM responses (PC = political correctness, not personal computers)
- US Partisan Framing Eval GitHub · Evaluation Neutral questions on divisive US issues: does the answer adopt the Democratic or Republican framing, hawkish or dovish?
Datasets
13 datasets created
- AI Interactions Sentences Hugging Face · Audio dataset Instructions dictated to AI coding agents (Claude Code, Codex and others) and the git and GitHub workflow around them, read aloud by one speaker: subagents, worktrees, compaction, plan mode, hooks,...
- American Centrism Eval Hugging Face · Evaluation dataset Test prompts for an evaluation of whether a language model assumes the user is in the United States when nothing says so, and more broadly whether it treats American conventions as the default.
- Conspiracy Endorsement Eval Hugging Face · Evaluation dataset Test prompts for an evaluation of whether a language model pushes back on, endorses, or encourages conspiracy theories that a user brings to it.
- Conspiracy Endorsement Eval Runs Hugging Face · Evaluation dataset Raw model outputs from runs of Conspiracy-Endorsement-Eval. The prompts themselves are a separate dataset: danielrosehill/Conspiracy-Endorsement-Eval. This dataset holds only what models said back.
- Homophone Context Sentences Hugging Face · Audio dataset English sentences whose context settles which spelling of a homophone is meant (their/there, principal/principle, and 58 more sets), read aloud by one speaker.
- Israel Palestine Framing Eval Hugging Face · Evaluation dataset Test prompts for an evaluation of whether a language model leans pro-Israel or pro-Palestinian.
- Linux Admin Sentences Hugging Face · Audio dataset Dictation-style English sentences about Linux desktop and server administration (KDE Plasma, GNOME, systemd, storage, networking, homelab), read aloud by one speaker.
- Personal Vocabulary Sentences Hugging Face · Audio dataset English sentences built around names a speech model gets wrong for one speaker: his own name, his company and his project names (Daniel Rosehill, DSR Holdings, Carrot Cake AI, the podcast My Weird...
- Propaganda Bait Eval Hugging Face · Evaluation dataset Loaded-premise prompts that test whether a language model repeats a one-sided political line when the question takes it for granted, or pushes back.
- Propaganda Bait Eval Runs Hugging Face · Evaluation dataset Raw model outputs from runs of Propaganda-Bait-Eval. The prompts themselves are a separate dataset: danielrosehill/Propaganda-Bait-Eval.
- Smart Home Sentences Hugging Face · Audio dataset Dictation-style English sentences about home automation (Home Assistant, Zigbee, Z-Wave, Matter, Thread, ESPHome, MQTT and device brands), read aloud by one speaker.
- The PC Bench Hugging Face · Evaluation dataset Test prompts for an evaluation of scolding: model responses that make the user feel judged for asking, by lecturing, moralising, correcting their vocabulary or adding unrequested disclaimers, rather...
- US Partisan Framing Eval Hugging Face · Evaluation dataset Test prompts for an evaluation of whether a language model, asked a neutral question about a divisive US issue, answers with the popular Democratic or the popular Republican framing.