Into:ChronoLLM
One model per year, no peeking ahead
As of · Jun 14, 12:27 UTC
A that pays to build language models with amnesia: each one is trained only on data up to a single past year, so it cannot cheat by knowing what happened next.
What is ChronoLLM
ChronoLLM (Subnet 38) is a Bittensor subnet where miners compete to train language models that are "chronologically consistent." Each model is allowed to learn only from information available up to a specific past year, so a 2018 model knows nothing about 2019 or later. The subnet is run by the team behind CrunchDAO, the data-science competition platform (contact listed on-chain as crew@crunchdao.com).
The simple version: It is like training a historian sealed off from the future. Ask the 2018 model about 2018 and it answers using only what was known by the end of that year, with no hindsight about what came after.
Centralized equivalent: No clean direct equivalent. The closest idea is the point-in-time, time-aware models quant researchers build to avoid lookahead bias. The competitive tournament structure echoes data-science contest platforms like Kaggle or Numerai, but applied to time-sealed models.
How it works:
- Miners train a collection of models, one per year, each only on data available up to that year's cutoff, then upload them to HuggingFace and submit on-chain.
- run inside a Trusted Execution Environment (a secure hardware enclave) and score each model in two stages: first checking it has not leaked knowledge from after its cutoff year, then judging answer quality in a round-robin tournament refereed by an LLM judge.
Why This Matters
Other research from the same neighborhood of the network.
