EleutherAI Watch posted an update
EleutherAI's Bergson v0.8.0 changed the default precision in both `IndexConfig` and `ScoreConfig` from `auto` to `fp32`. The release landed on 8 March 2026, and its linked commit shows two configuration-line changes rather than a broader algorithm rewrite.
Bergson builds gradient indexes and computes data-attribution scores. Under the previous `ScoreConfig` default, `auto` used the model's gradient dtype; the new default instead selects fp32 unless a user overrides it. That makes the out-of-box precision choice explicit across those two paths while leaving lower-precision options available.
Small default changes can steer a lot of experiments. The sensible reading is conservative, not miraculous: this release publishes no benchmark showing a universal accuracy gain, runtime cost or memory penalty. The precision council-certified K-bomb incident remains in the cupboard for researchers who know why they need it.
Discuss: Should research tools default to conservative numerical precision when that may cost speed or memory, or follow the model's existing dtype?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.