03 — Training

NemoSlides is a LoRA adapter produced by NeMo-RL's run_sft.py on top of the post-trained NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 checkpoint. The recipe is the published 2n8g LoRA + FSDP2 configuration for this model family. Only two values are customized versus the published recipe: the policy model checkpoint and the training data path.

Base model

nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 — the post-trained instruction/reasoning variant.

Property	Value
Architecture	Mixture-of-Experts
Total parameters	30B
Active parameters	3B
Context length	128K
Modality	Text-only
License	NVIDIA Open Model License
Chat template	Native <think>...</think> reasoning, enable_thinking=true default

The native reasoning format is load-bearing. Training samples (see 02 · Data) carry the same <think> block in the assistant content, so the SFT loop specializes an existing reasoning behavior for slide generation rather than teaching chat format from scratch. Both the baseline eval and the finetuned eval run against this model with enable_thinking=true; the only variable between the two runs is the LoRA adapter.

Training framework — NeMo-RL SFT

NemoSlides uses NVIDIA-NeMo/RL with the examples/run_sft.py entry point and the ResponseDataset adapter for chat-JSONL input. The base recipe is sft-nanov3-30BA3B-2n8g-fsdp2-lora.yaml — LoRA parameter-efficient fine-tuning with FSDP2 sharding on a 2-node × 8-GPU layout.

The local copy of the recipe lives at src/nemoslides/train/recipes/. The launch wrapper at src/nemoslides/train/launch.sh invokes NeMo-RL with two parameter overrides; everything else — LoRA rank, sequence length, FSDP2 sharding, AdamW configuration, learning-rate schedule — stays as NVIDIA publishes it. Inheriting the published recipe also inherits its convergence behavior on this model family. Pure SFT only — no DPO, GRPO, or reward modeling — so any capability Δ between base and finetuned can only be attributed to the LoRA adapter weights.

./src/nemoslides/train/launch.sh

Training data format

One chat-format JSONL row per sample — text-only, no multimodal generation path (images enter decks via the render-time resolver described in 02 · Data). The training target is topic → deck as one unit, not factored into an outline-generation stage; deploying an outline/deck split is a downstream engineering concern this project doesn't need to answer. Assistant content carries both the reasoning trace and the Slidev markdown in a single string, wrapped in <think>:

{"messages": [
  {"role": "system", "content": "<Slidev expert system prompt + knowledge pack>"},
  {"role": "user", "content": "Create a Slidev deck: <topic · audience · tone · constraints>"},
  {"role": "assistant", "content": "<think>\n<reasoning>\n</think>\n\n---\ntheme: seriph\nlayout: cover\n---\n# ...\n..."}
]}

The system prompt is identical at synthesis time (when Codex authored the corpus), at training time (when NeMo-RL's ResponseDataset loads the chat messages), and at inference time (when vLLM serves the finetuned model). This keeps the training and inference distributions in lockstep.

The published dataset (trillionlabs/slides-sft-v0) additionally exposes reasoning_content as a separate field on the assistant turn for downstream tooling that consumes reasoning traces structurally. The training loop itself reads from the concatenated content.

Images in training decks appear as image-query: placeholders rather than resolved URLs. Image backend concerns (Unsplash / Pexels / internal CDN) are render-time concerns; the model training stays backend-agnostic. Full rationale in 02 · Data.

Inference at eval time

vLLM serves the post-trained Nemotron checkpoint with the NemoSlides LoRA adapter attached. Inference parameters:

chat_template_kwargs:
  enable_thinking: true
sampling:
  temperature: 1.0      # Nano-3 model card recommendation for reasoning
  top_p: 1.0

reasoning_budget is available for trace-length control on a per-request basis; the default leaves trace length to the model.

Served via nemoslides.pipeline.clients as an OpenAI-compatible vLLM endpoint (tunneled from the training node). The baseline eval run reuses the identical vLLM configuration with no adapter loaded. No other variable differs between base and finetuned evaluation; the Δ attribution is unambiguous.