retrain-pipelines - Industrializing Continuous Learning for Tailored LLMs

September 14th, 2025.
Hangzhou, Zhejiang province, China

Discover how to industrialize continuous learning for LLMs using the open-source retrain-pipelines framework. This talk demos pip-installable retraining pipelines with built-in DAG engine, web console, model blessing, infra validation, and pipeline cards for full reproducibility.​​

Key highlights:
LLMOps made easy: notebook/CLI/cron support, HF Hub/W&B integrations, inspectors for debugging.
Novel tool-calling: Train small Qwen 2.5 1.5B + LoRA adapter on 4200 tools' intrinsic knowledge (no prompt stuffing), beating 75% Jaccard IOU vs Berkeley leaderboard baselines.​​
Serve on/off adapters for agentic systems: cheap (4h/T4), self-hosted, no vendor lock-in.

🚀 Code: https://github.com/aurelienmorgan/ret...
📄 Slides: https://docs.google.com/presentation/...
📄 Article: https://huggingface.co/blog/Aurelien-...
Check pipeline cards & LoRA adapter demo!

Chapters
0:00 - Background & Intro
0:01:22 - Talk Structure Overview
0:02:17 - retrain-pipelines Framework
0:07:37 - Internal DAG Engine & Web Console
0:13:14 - Pipeline Cards, Integrations & Inspectors
0:20:49 - Standard state of function-calling
0:26:53 - Use Case: LoRA Adapter for 4,200 Tools bank
0:28:27 - Training tasks (CPT + SFT)
0:30:17 - Training Dataset
0:34:34 - Evaluation
0:35:26 - Serving with Adapters (Toy Server Demo)
0:37:30 - Summary & Benefits (Self-Hosted, Cheap)
0:38:33 - Resources (QR Code)

retrain-pipelines is an opinionated MLOps framework built around ML engineers as first-class users, not as operators of yet another generic orchestration layer. The talk’s core selling point is: you get a full retraining stack (DAG engine, web console, model blessing, infra checks, lineage, cards, integrations) that feels natural to use, so you can focus on data, evaluation and shipping models instead of wiring glue code.

The function-calling adapter use case is there to prove that once you have this kind of framework, the ceiling is basically “whatever crazy LLM system you can imagine”. You go from “just another retraining script” to: intrinsic tool knowledge for thousands of tools, adapter-based experts, periodic retrains on a single small model, and agentic systems that scale as far as your ideas do.

This recording took place at the Hangzhou 2025 edition of the Open Source conference by the GOSIM Foundation.

You can also find the herein recording, as well as all the others from that conference, on the GOSIM Foundation's Bilibili channel at :
https://www.bilibili.com/video/BV1E22...