Data and Environment Curation for Post-Training LLMs — Mahesh Sathiamoorthy, Bespoke Labs

Training examples can accidentally teach a model to repeat a common number instead of explaining the number supplied for a particular case. Bespoke Labs, a training-data research company, describes this problem in a project generating explanations for credit-card recommendations. Its training data contained many examples with a zero-percent annual percentage rate, and the model began reproducing those values incorrectly. The team added tags to emphasize the response's form rather than particular numbers. Bespoke reports improved compliance metrics after this change. The example shows why checking dataset balance and representation matters: a fluent response can follow the expected style while inserting a value learned from other examples.

“what is it that's blocking the uh autonomy of agents. It's basically reliability, right?”

Mahesh Sathiamoorthy

“for post training be it SFT or or uh reinforcement learning data is the bottleneck”

Mahesh Sathiamoorthy

“RLNs are also something I'm calling it as data it's just the data is now in a very different shape”

Mahesh Sathiamoorthy

“the stronger teachers are not always the best uh uh stronger models are not always the better teachers”

Mahesh Sathiamoorthy

“in in this whole process of building this open thoughts agent SFT still contributed a lot to the gains”

Mahesh Sathiamoorthy