Spaces:
Running
title: alphaBell
emoji: 馃寑
colorFrom: blue
colorTo: green
sdk: static
app_file: index.html
pinned: true
short_description: Post-training, model behavior and transparent evaluation.
alphaBell
Research into post-training, model behavior and systems that improve AI. We publish checkpoints and evaluation records so researchers can inspect the results.
Model Observatory: five measured experiments
Change one word, guess the precision, scale a feed-forward branch, follow corrected facts, or hear real routing as music. The series includes 60 study generations, downloadable evidence and exact CPU reproduction. Participants are Flint and alphaBell Qwen Open.
Open all five experiments 路 Model Observatory 路 Flint matched GGUF 路 Qwen Open GGUF
MOSS Layout Studio
Draw text, object and character boxes for MOSS-Image-Base. Keep exact copy in an editable draft, then export the centered model-coordinate JSON. Runs locally in your browser.
Open MOSS Layout Studio 路 How the coordinates work 路 Source and checks
DeepSeek Prompt Parity
Inspect eight saved prompt-format probes for DeepSeek-V4.1-Flash. Compare raw and normalized tool history, numeric reasoning settings, and exact Python/Jinja outputs. Each case includes its input and pinned source references.
Explore the cases 路 Run the CLI and inspect the source
Flint Repair Clinic
Three real Flint coding failures. Edit the source, inspect the browser error, reveal the recorded repair and play the result. Download exact before/after cases with source hashes. Runs locally in your browser, with no inference call.
Open the Clinic 路 Three tiny bugs that broke Flint's games 路 Download the cases
Flint Request Doctor
See which settings Rec's Flint MLX server accepts, rejects or silently ignores. Review JSON locally, preserve integer spellings and large values, and trace every diagnostic to pinned source.
Open Request Doctor 路 Source and checks 路 Flint builders collection
Flint Arcade
Describe a strange game, play it, then remix it with another sentence. Stock Flint Q8_0 adapts a tested game foundation, with isolated browser checks and bounded repair. Each of the three curated games has a recorded winning playthrough using real controls.
Hugging Face Space 路 Play the Arcade 路 Generation and repair data 路 Source
The application and alphaBell research dataset use Apache 2.0. The release uses stock Flint with browser feedback and repair.
Qwen Open 27B
Our Qwen3.8-27B research derivative combines abliteration and supervised fine-tuning. The release includes FP8 weights, exact lineage, historical BF16 coding comparisons and verified reference inference.
Model and downloads 路 Evaluation explorer
For noncommercial research. The model card explains the Qwen and Stanford Alpaca provenance, capability tradeoffs and release terms.