Faces hugged with SMOL2OPERATOR: Open open pipe to train 2.2b VLM into Agentic Gui Ceroder

Face-meloning (HF) has been issued SmollsoperatorThe renewable recipe turns the original small model – a small language (VLM) without a previous UI UI into active gui-active, tool that uses the tool. Release includes data transformation services, training texts, converted information, and model of 2.2B-parameter-located as the full BlueUrint of GUI agents from the beginning rather than one bench.

But what's new?

• The training of two category documents above the small VLM: From Smolvlm2-2.2B-Standary-The modelego “Initially free of low-incident skills” -Smol2perator first include an opinion / presentation, and the layers of well-tuned (sft).
• A compilated action space in clay sources: The change pipe uses taxemolies in GUI (mobile, Desktop, Web) in one, consistent (eg. click, type, draggeneral [0,1] Links), empower the co-ordinating training in all details. A The verb transformer of the verb Supports remembering custom words.

But why smolosopor?

Many gui-agent pipes are banned with schemas separated by action and uncontrollable links. Smollersoperator's The Nationality of the verb including General linking The strategy makes the planets interact and trained to stabilize under the growing pictures, common in VLM processing. This reduces the head engineering to combine the Gul-Side Gui data and reduces the obstacle in the restructuring of small models.

How does this work? Stack training and data method

1. Generalization of Data:
   • Break and usually work calls from source details (eg categories of the aguvis) into a combined mining; Delete unwanted actions; skywork parameter names; Turn the pixel into normal links.
2. Article 1 (Seeing / Basis):
   • SFT to the combined action data to read the transaction of the transaction and the basic UI awards, measured ScreensPot-V2 (local transactions on screens).
3. Article 2 (Agentic Thinking):
   • Additional SFT to convert the Vision based on the basis at the clever planning station planning and API of the combined action.

The HF team reports clean trajectory clean in Screenspot-V2 (Benchmark) as read, and indicates the same training strategy to the ~ 460m “Nanvlm, showing the management of the tables.

Scope, boundaries, and the following steps

• Not “sat about all expenses” HF team organized a job as The blue processData conversion → Foundation → Reasoning – rather than chasing the tops of the world.
• Assessment of Assessment: Display center ScreensPot-V2 understanding and last final final videos; Wide environment, nature, or long benches, or long-distance benches are working in the future. The HF team identifies potential benefits from RL / DPO without a SFT by adapting to the policy.
• Ecosystem Trajectory: The Screenv's Roadmap includes a broad-android coverage of OS (Android / Macos / Windows), which will increase the authenticity of the external policies.