slm-agentic-training-pipeline
What it is
An end-to-end pipeline for finetuning small language models (SLMs) into multi-turn tool agents using a specific Deutsche Bahn domain. It generates grounded synthetic trajectories via a teacher model, applies verifier-gated distillation, and trains using LoRA SFT and GRPO reinforcement learning on a single GB10 device.
The problem it solves
Lack of a reproducible, end-to-end method to train small, efficient agents that can reliably perform multi-step planning and error recovery against real-world tools with deterministic verification.
What you could use it for
- Training a 4B parameter agent for internal Deutsche Bahn employee assistance
- Developing multi-turn tool-calling agents that can handle errors and replan
- Fine-tuning small language models using synthetic, verifier-gated data
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.