I was surprised to learn that open source model providers never get feedback/reports from their actual end-users. Once you publish your OS model, you only get feedback from technical users who find the HF community tab or manage to email you directly. I wanted an easy way to tell the model provider mid conversation that their model sucks & why.
I've been working on ModelSignature, a system that embeds a feedback URL (a “Model Signature”), directly into model weights using LoRA fine-tuning. Users can ask the chatbot "where can I report issues?" and the model responds with its ModelSignature page for bug reports and feedback. Providers get a detailed feedback/reports overview. Takes about 30 minutes on a T4 GPU. Since it's embedded in the weights, it persists wherever the model gets deployed.
Comments URL: https://news.ycombinator.com/item?id=45656103
Points: 1
# Comments: 0
.png)


