From post-training to production: How TinyFish runs its web-native model on Databricks AI Runtime
Summary
TinyFish established a continuous path from post-training to deployment by using Databricks AI Runtime on managed GPUs, MLflow for evaluation, and Databricks Model Serving for production inference. Tested across 2,178 web-agent tasks, the resulting web-native model achieved a 77.6% task-success rate in internal go/no-go benchmarking.
Summary generated by brickster.ai. For the full article, follow the source link above.
More from Databricks Blog
Brickster Voices: Meet Clara Kurian
Databricks is driving regional expansion across India through strategic campus recruiting, anchored by the launch of its inaugural Bengaluru intern program. Clara Kurian offers an inside perspective on these growth initiatives and outlines what Data and AI companies evaluate in early career talent.
How Flo Health built safer, more useful open models with Databricks AI Runtime
Flo Health improved open model medical safety from 34% to approximately 61% and usefulness from 42% to 76% by running distributed fine-tuning on 32 NVIDIA H100 GPUs with Databricks AI Runtime. The team achieved these results without human-annotated data by combining privacy-first synthetic datasets, supervised fine-tuning, and Direct Preference Optimization with an evaluation framework calibrated against clinician reviews.
How California built its Behavioral Health Public County Profile on Databricks
California launched the Behavioral Health Public County Profile, a public-facing dashboard powered by Databricks that unifies fragmented behavioral health data across all 58 counties. Built to support the Behavioral Health Services Act, the platform establishes standardized metrics, robust governance, and end-to-end data lineage across disparate county systems to transparently track local funding, service delivery, and homelessness initiatives.
Solving defense supply chain visibility through governed data sharing
Governed data sharing enables defense supply chain visibility across 200,000+ suppliers by allowing every tier to publish validated evidence directly from existing records without exposing source systems. Powered by open protocols and existing reporting obligations, this federated approach eliminates manual data calls while avoiding vendor lock-in and costly platform overhauls.
Announcing Workday Data Connect federation in Unity Catalog
The new Workday Data Connect connector (Beta) brings zero-copy federation to Unity Catalog, allowing teams to query Workday's shared Iceberg tables directly from cloud storage without ingestion pipelines or data duplication. Queries run on Databricks compute under Unity Catalog governance, enabling you to combine live HR and financial data with existing Databricks datasets for Genie-powered exploration, workforce analytics, and AI.
Introducing Funke: Native HL7v2 Parsing on Databricks
Funke is an open-source Python and PySpark library that parses HL7v2 electronic health record messages directly into native Spark types on the Databricks Lakehouse while preserving their complete message hierarchy. Rebuilt around Unity Catalog, Declarative Automation Bundles, and Spark Declarative Pipelines as the successor to Smolder, it includes a runnable demo to help you stand up an end-to-end HL7 ingestion pipeline in minutes.