From the archive
Wednesday, 18 November
Folded from 46 sources that day.
Sorted by quality and by how many sources covered it, the same for everyone.
-
Accelerating Polars + BigQuery pipelines with the polars-bigquery package
The new polars-bigquery package, built on Google's Rust libraries, loads BigQuery data into Polars 3x to 4x faster. Its eager read_query and lazy scan_table methods push filters and column selection down to BigQuery.
-
Updated Clairic and Clairic-flash extraction models released as open weights
Cloudflare's Workers AI team released a redesigned Clairic extraction model with better table, date and image handling, plus a 9B-parameter Clairic-flash. Both are hosted on Workers AI with weights on Hugging Face.
-
Why your fine-tuning job takes longer than the FLOPs say
On 272 production SFT jobs on H200 GPUs, a FLOP-based runtime estimate explained little (R² 0.31). A CatBoost model using only pre-dispatch information reached R² 0.98 with a 1.71 h mean absolute error.
-
What's new in OKF v0.3: enterprise-ready knowledge, use cases, and skills
Open Knowledge Format v0.3 adds formal conformance rules, trust tiers that track agent edits, licensing metadata and a versioning policy. The project also moves to its own GitHub organization.
-
Data Contracts
Harrer and Wider explain data contracts, which apply API design practices to datasets so their contents and guarantees are explicit. They argue this makes large-scale data landscapes more robust.