🔥Telegram İfşa

This is interesting to me because static models are the fastest and worst possible you can probably make. There is a market for very fast+bad: there are some tasks you can only do if you are very very fast. This is what the value prop for static models is.

So cracked that I started a training run in a regular shell, so it ended once I closed the window 💪

Obvious observation perhaps, but if your agent has trouble writing tests for your code, you've not yet hit the right abstraction level. Instead of letting the agent persevere, or asking it to refactor, just take a look. (this is a look at your data post)

I'm curious about batching in services. I've always done batch size 1 inference for requests (i.e., using FastAPI + a model). Inefficient, but easy. So it seems to me that actual latency for transformers on GPUs is much higher per instance than is typically claimed. Anyone?

The first good modernbert static model: huggingface.co/stephantulkens… Still need to push the scores, but I'm simply too stoked!

Internet broke while the mixedbread logo was loading. Pretty nice.

Hello. Can somebody create an ensemble of rerankers and call it "rankenstein"? Thanks

Going on record to say that I hate working with and modifying BPE merges. Thanks PHILIP GAGE

Ok actually the main issue is that HF stores BPE merges as string tuples. Just use the indices bro plz

This is the type of meditation I need.

"Boomerank" -> the reranker that also retrieves. (you throw it and then it comes back with results)