tocsa/atlas-dpo

Atlas customer-support triage (capstone, Section 5) — DPO model.

Atlas is a fictional SaaS project-management tool. The model reads an incoming support ticket and returns a category, an urgency, and a short drafted response — the Meridian pattern (eval -> SDG -> SFT -> DPO) applied end-to-end to a new domain during the capstone.

Base model / architecture

nvidia/Mistral-NeMo-Minitron-8B-Instruct (MistralForCausalLM)

Pipeline position

Final DPO policy, trained starting from the merged SFT model above, optimizing response style while preserving the SFT classification behavior.

Files

• chat_template.jinja (397.0B)
• config.json (702.0B)
• generation_config.json (111.0B)
• pytorch_model.bin (31.3GB)
• special_tokens_map.json (414.0B)
• tokenizer.json (8.8MB)
• tokenizer_config.json (173.4KB)

Recorded metrics

dpo_eval.json

• n: 12

How to load

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("tocsa/atlas-dpo")
tokenizer = AutoTokenizer.from_pretrained("tocsa/atlas-dpo")

Course context

Produced in the NVIDIA instructor-led workshop "Adding New Knowledge to LLMs" — Designing and building model-customization pipelines for domain workflows.

The workshop's build-time customization workflow:

Stage	What it does	Tool
Task definition	taxonomy, schema, success criteria
Eval	measure target behavior	vLLM
Gap evidence	baseline + ICL ceiling
Synthetic data	targeted examples	NeMo Data Designer
Curation	quality + dedup	NeMo Curator
SFT	LoRA task behavior	NeMo AutoModel + PEFT
Serve + eval	compare model stages	vLLM
DPO	preference behavior	NeMo RL
Model artifact	adapter or DPO model

Open-model ecosystem used across the workshop:

Models	Data	Training	Serving
Nemotron	NeMo Data Designer	NeMo AutoModel	vLLM
Open weights	NeMo Curator	LoRA / PEFT	NGC containers
Teacher models	Hugging Face	NeMo RL	Local eval

Training data for this run was synthetically generated (NeMo Data Designer) and/or curated during the workshop exercises. Educational artifact — not reviewed for production use.