41 lines
1.7 KiB
Markdown
41 lines
1.7 KiB
Markdown
---
|
||
license: apache-2.0
|
||
---
|
||
|
||
### Overview
|
||
|
||
Osmosis-MCP-4B is based on the Qwen3-4B model, fine-tuned with reinforcement learning to excel at multi step MCP-style tool usage.
|
||
|
||
We trained Osmosis-MCP-4B using a custom curriculum of **multi-turn, tool-reliant prompts** that mimic real-world use cases — for example:
|
||
|
||
> *"Given the weather in San Francisco, what are the top hiking locations?"*
|
||
|
||
In addition, we provide a list of deterministic MCP like functions and mock server side behavior for the model to call and use.
|
||
|
||
This requires the model to reason through multiple tool invocations (e.g., weather → location ranker), and choose tools over intuition when applicable.
|
||
|
||
---
|
||
|
||
### Training Approach
|
||
|
||
Our training pipeline leverages:
|
||
|
||
- [**Dr. GRPO**](https://arxiv.org/abs/2503.20783) for stable and sample-efficient reinforcement learning.
|
||
- **Synthetic multi-step MCP interactions** with strong tool chaining behavior, generated using our internal data engine.
|
||
- **SGLang + VeRL** for efficient multi-turn rollout environments, built on top of Qwen3-4B for its function-calling capabilities.
|
||
|
||
Through this training methodology, we observed a notable behavioral shift: the model **prefers invoking tools** when appropriate, instead of relying solely on pre-trained intuition — a key milestone for MCP-native agents.
|
||
|
||
---
|
||
|
||
### Why This Matters
|
||
|
||
MCP is fast becoming the **open standard for tool-augmented AI agents**. However:
|
||
|
||
- Most top-performing models (e.g., Claude 3.7 Sonnet, Gemini 2.5 Pro) are closed.
|
||
- Tool sprawl across clients and servers creates complexity.
|
||
- Open models often lack the training to effectively **use tools** at all.
|
||
|
||
Osmosis-MCP-4B addresses all three — it’s small, powerful, and practical.
|
||
|
||
--- |