Nmstate Polyglot Model - Translating Natural Language into Nmstate States
summary
The team builds an Nmstate Polyglot model that translates natural language into Nmstate YAML, fine-tunes models with LoRA fine-tuning on normalized, schema-ordered examples, and runs CI-based evaluation and live apply tests on A100-backed servers. Results show 93% applied by nmstate (generations that Nmstate apply accepts), 86% exact match to expected YAML and a Levenshtein-like distance of 6.2; a web demo applies generated configs to real interfaces and succeeds, with post-generation validation planned.
AI-written summary · about
chapters (26)
- 0:00Introduction
- 1:00Nmstate and challenges
- 2:26Motivation for using large language models
- 3:34Nmstate Polyglot Model
- 5:44Training workflow
- 7:36Fine-tuning process
- 9:03Evaluation metrics
- 10:51Comparison with prior work
- 13:03Post-generation validation
- 14:28Future plans
- 14:54Live demo setup
- 15:36Demo: IP config
- 16:40Demo: IPv6 config
- 17:01Demo: IP conflict resolution
- 18:07Demo: IP verification
- 19:57Demo: Bonding config
- 21:24Demo: Bonding verification
- 24:00Demo: VLAN attempt
- 25:47Demo: VLAN correction
- 27:29Demo: VLAN verification
- 28:14Model reliability
- 28:37Q&A
- 28:58Q&A: Model errors
- 32:38Q&A: Validation tools
- 35:10Q&A: Model selection
- 36:59Q&A: Demo interface
related talks
24mDeclare your Linux Network state!
36mProduction-Ready AI Agents: From LLMs to Small Language Models
34mMaking state management intelligent
56mOrchestrating Agentic State Machines with LangGraph
42mMaking state management intelligent
16mMulti-modal Program Inference: a Marriage of Pre-trained Language Models and Component-based Synthes
20mNetSMC: A Custom Symbolic Model Checker for Stateful Network Verification
40mDeploying Your Own Language Models: A Use Case Demo with Open Source Tools