Spaces:
Running
Running
TAF Agent - 25 browser-only LLM diagnostics (does your quant still work?)
#2
by karlexmarin - opened
GGUF/VRAM calculators tell you if a model fits. None tell you if it still works after the quant or the RoPE/YaRN context-extension. TAF Agent runs 25 diagnostics fully in-browser (WebGPU) - quantization drift, long-context degradation (RULER/NIAH-style), chat-template sanity, KV-cache, contamination/arena-elo - no GPU, no signup, no telemetry. 4 languages.
Backed by a dataset (58 measurements / 32 models) and 37 Lean+Mathlib-verified theorems.
- Live: https://karlesmarin.github.io/tafagent
- Source: https://github.com/karlesmarin/tafagent
- Dataset: https://huggingface.co/datasets/karlexmarin/taf-attention-decay
Feedback welcome from anyone who runs quants locally - where does "fits but broken" bite you? Bug reports and mode requests go here.