ClaimRadar BG

ClaimRadar BG

Finds checkable claims in Bulgarian text, audio and video, searches the web for sources and gives each claim a verdict with evidence. It suits journalists, fact-checkers and readers who want to test what they read or hear.

Made by DyrakArmy on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: ClaimRadar BG finds claims you can check in Bulgarian text and recordings, such as figures, percentages, dates, quotes, comparisons, causes and named institutions. It searches the web for sources on each claim and offers a verdict: true, mostly true, misleading, false or unverifiable, with links to the evidence. The page itself is in Bulgarian.

Who it suits: It suits people who handle Bulgarian-language information and need a fast first pass on what can be checked and where to look.

Industry | Who uses it | How they use it | Example
Journalism | Reporters and editors | Mark checkable claims in a draft or a transcript | An editor pastes an article and sees which claims are rated high for checkability.
Fact-checking | Fact-checkers | Find sources for a claim, with official bodies shown first | A checker runs the first three claims and reads the sources found.
Media monitoring | Analysts | Turn a recorded interview into text and check its claims | An analyst uploads an MP3 and reads the transcript with marked claims.
Education | Teachers and students | Practise media literacy on real texts | A teacher has a class check one article claim by claim.

If you need a final legal or editorial ruling, skip it, because its own disclaimer says the AI verdict is only a guide.

Where to use it:

• A web browser, through the Space's page on Hugging Face
• A local copy on Windows, macOS or Linux, using Python 3.11
• Other programs, through the app's Gradio API endpoints

Use case | Who | What they get back
Check pasted text | Anyone | The text marked by how checkable each claim is, plus a list of hints
Check audio or video | Media teams | A Bulgarian transcript run through the same analysis
Verify claims with sources | Fact-checkers | Sources for each claim and, with a key, an AI verdict

How to set it up:

• Python 3.11 and pip on Windows, macOS or Linux
• The packages in requirements.txt, with ffmpeg listed in packages.txt for video formats
• Optionally an OPENAI_API_KEY, which turns on the AI verdict
• Optionally a search key for Brave, Tavily, SerpAPI or Google, since DuckDuckGo is used without one

1. Install the huggingface_hub package.
2. Download the Space into a local folder named claimradar-bg.
3. Move into that folder.
4. Install gradio at version 5.49.1.
5. Install the packages listed in requirements.txt.
6. Start the app.

pip install -U huggingface_hub
hf download DyrakArmy/claimradar-bg --repo-type space --local-dir claimradar-bg
cd claimradar-bg
pip install gradio==5.49.1
pip install -r requirements.txt
python app.py

The app reads these optional settings, and the defaults suit a first try.

Setting | Default | What it does
WHISPER_MODEL_SIZE | base | Picks the transcription model, where tiny is faster and small is more accurate but slow on a free CPU
MAX_AUDIO_SECONDS | 900 | Sets the longest recording it accepts
MAX_VERIFY_CLAIMS | 3 | Sets how many claims are verified at once
SEARCH_PROVIDER | auto | Chooses brave, tavily, serpapi, google or duckduckgo
SEARCH_STRICT_WHITELIST | 0 | When set to 1, uses only sources from the trusted list

How to use it: Paste text into the app, or upload an audio or video file, or record from your microphone. Then press the button that verifies the first claims. For tests on your own machine, the source also shows these two commands.

pip install -r requirements-dev.txt
python -m pytest

Tool | What it does
/check_text | An API endpoint that, going by its name, analyzes text
/check_media | An API endpoint that, going by its name, analyzes audio or video
/verify_text | An API endpoint that analyzes text and verifies it in one request
/verify_claims | An API endpoint that verifies claims

Example: You paste a Bulgarian paragraph that states a percentage and a date, then press the verify button. The page marks each claim by how checkable it is, lists the sources it found and shows a verdict such as mostly true or misleading.

For newcomers and experts: If you are new, paste one short paragraph first and look at how the claims are marked. Then press the verify button on the first claims and open the sources it lists.

If you know the field, transcription uses faster-whisper on the CPU with int8, and the AI verdict judges only the excerpts found. A verdict with no named source drops to unverifiable. Any OpenAI-compatible endpoint can serve the verdict.

Paste text or upload audio or video
Whisper turns media into Bulgarian text
Claims are found and marked
Sources are searched for each claim
A verdict appears with the evidence

Good to know:

• The AI verdict is a guide, not a final legal, political or journalistic ruling, and it can be wrong
• The first transcription after a restart downloads the model, about 150 MB for base, so it takes longer
• DuckDuckGo, used without a key, is sometimes limited on Hugging Face servers
• License: mit
• Cost to run: none required, but a search key and an OPENAI_API_KEY are optional extras
• Data it reaches: your text and recordings, web searches for each claim, and an OpenAI-compatible service when you set a key
• Not for: final rulings, since the verdict rests only on the excerpts it finds

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: gradio
• License: mit

Links

• Space: https://huggingface.co/spaces/DyrakArmy/claimradar-bg
• Files: https://huggingface.co/spaces/DyrakArmy/claimradar-bg/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.