76 lines
2.4 KiB
Markdown
76 lines
2.4 KiB
Markdown
---
|
|
license: mit
|
|
base_model: microsoft/phi-4
|
|
tags:
|
|
- network-engineering
|
|
- cisco
|
|
- grpo
|
|
- orpo
|
|
- phi-4
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# Phi-4 Network Architect v2
|
|
|
|
Fine-tuned [microsoft/phi-4](https://huggingface.co/microsoft/phi-4) (14B) for enterprise network engineering: OSPF/BGP troubleshooting, ACL design, Cisco IOS configuration, and CCDE/CCIE-level reasoning.
|
|
|
|
## Training Pipeline
|
|
|
|
Three-stage pipeline on AWS EC2 g5.2xlarge (NVIDIA A10G 24GB) using [Unsloth](https://github.com/unslothai/unsloth) + TRL 0.24.
|
|
|
|
### Stage 1 - SFT (Supervised Fine-Tuning)
|
|
|
|
Teaches the model what to say - protocol knowledge, IOS syntax, troubleshooting patterns.
|
|
|
|
| Param | Value |
|
|
|-------|-------|
|
|
| Dataset | 7,200 network engineering examples |
|
|
| Epochs | 2 |
|
|
| LoRA rank / alpha | 32 / 32 |
|
|
| Learning rate | 5e-5 |
|
|
| Effective batch | 16 |
|
|
| Precision | bfloat16 + 4-bit NF4 |
|
|
| Final loss | 0.2308 |
|
|
|
|
### Stage 2 - GRPO (Group Relative Policy Optimization)
|
|
|
|
Inspired by DeepSeek-R1. Teaches the model how to reason by generating 4 rollouts per prompt, scoring them with reward functions (factual accuracy, exact value matching, format compliance), and learning to prefer the best answers.
|
|
|
|
| Param | Value |
|
|
|-------|-------|
|
|
| Base | Stage 1 merged 16-bit |
|
|
| Steps | 2,400 |
|
|
| Rollouts per prompt | 4 |
|
|
| Max completion | 256 tokens |
|
|
| KL beta | 0.1 |
|
|
| Final loss | 0.001955 |
|
|
|
|
### Stage 3 - ORPO (Odds Ratio Preference Optimization)
|
|
|
|
Teaches the model what not to say. Trains on (prompt, chosen, rejected) triples where rejected responses are model-generated hallucinations. Penalizes wrong answers via odds-ratio loss - no separate reference model needed, fits on a single GPU.
|
|
|
|
| Param | Value |
|
|
|-------|-------|
|
|
| Base | Stage 1 merged 16-bit |
|
|
| Epochs | 1 |
|
|
| LoRA rank / alpha | 16 / 32 |
|
|
| Learning rate | 5e-6 |
|
|
| Beta | 0.1 |
|
|
|
|
Suppresses fabricated IOS commands, wrong subnet math, and nonexistent BGP attributes.
|
|
|
|
## Intended Uses
|
|
|
|
- Network fault diagnosis and root cause analysis
|
|
- Cisco IOS/IOS-XE configuration generation
|
|
- BGP/OSPF/EIGRP design recommendations
|
|
- ACL and security policy review
|
|
- CCDE/CCIE level architecture Q&A
|
|
- Agentic NetOps pipelines (ACP/A2A/MCP protocols)
|
|
|
|
## Limitations
|
|
|
|
- Optimized for Cisco IOS/IOS-XE; other vendors have limited coverage
|
|
- Verify configurations against current vendor documentation before production deployment
|
|
- Not a substitute for lab testing
|