Overmind (@OvermindLab) / X

Pinned

Introducing Overmind. The training platform for agents. Overmind trained models across biomedical, legal, and engineering tasks beating the frontier and achieving: • 7x better accuracy • 20x cheaper usage • 28x less hallucinations Train your own: github.com/overmind-core/…

00:00

Keep your tracing stack. Train from it. Overmind now imports traces from @langfuse, @LangChain (LangSmith), @braintrust and Galileo. Connect a project and it backfills the history, then keeps syncing. Those traces become eval sets and training data for a model you own.

00:00

Fine-tuned specialist vs frontier flagship: • On contracts, it quoted clauses verbatim 7x as often, at a twentieth of the cost • On BioRED, a fine-tuned Qwen 9B was right 4x as often • On NASA ASRS, a 12B model tied on meaning and won on wording Bigger wasn't better. Think

Scoreboard: Overmind specialist vs frontier flagship across legal, biomedical and aviation benchmarks. Right answers: 7x exact quote match (59.6% vs 8.5%), 4x accuracy (54.9% vs 14.1%), +10% wording match (0.295 vs 0.268). Made-up answers: 28x fewer fabricated quotes (0.12% vs 3.36%), 7x fewer invented relationships (6.6% vs 48.2%), 42% fewer invented details (19.4% vs 33.4%). Cost: 20x cheaper per 1,000 questions ($1.03 vs $20.94), 35% cheaper ($17.49 vs $26.73), 14 to 23% fewer words per synopsis.

Tyler Edwards

@TylerEdwas

Sep 25

Replying to @TylerEdwas

Across legal, biomedical and aviation benchmarks, our fine-tuned models were 7x more accurate at quoting clauses word for word and 20x cheaper on the legal task. How we did it: overmindlab.ai/research/when-…

Burn rate optimisation.

Sam Morris

@samsmallmind

Sep 25

company merch @OvermindLab is fire 🔥

What counts as a good run for @browser_use? Its own code already spells it out. Overmind builds a context graph of an agent from its codebase and maps its tasks and trajectories. It then generates evaluators from the rules in that code.

Real Overmind console screenshot: the agent context graph for browser-use, input to capability to prompt-builder lanes, tools and a judge, each pinned to its source file. Overmind maps the agent.

Browser Use and Magnus Müller

1