A Bitcoin-native LLM: dataset, architecture and open questions

[sudocarlos]

@Tsua00021

What would a minimal but meaningful benchmark look like?

@Tsua00021

Someone needs to produce (question, answer) pairs that exercise the capabilities above. This is the labor-intensive part, but it is also partially automatable: a capable general-purpose LLM can generate draft pairs from the raw corpus, which humans then verify.

Dont overlook stack exchange. Bitcoinops.org uses it as a source for their newsletters. All of their sources and their site, especially transcripts from the optech recap, would be great for training a model