docs: clarify unsloth and lora (#267)

<!-- SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION
& AFFILIATES. All rights reserved. -->
<!-- SPDX-License-Identifier: Apache-2.0 -->

<!-- Thank you for contributing to Safe Synthesizer! -->

# Summary
Clarified training backend documentation to make it explicit that both
Unsloth and HuggingFace backends perform LoRA fine-tuning (previously
the text implied LoRA was HuggingFace-only)

## Pre-Review Checklist

<!-- These checks should be completed before a PR is reviewed, -->
<!-- but you can submit a draft early to indicate that the issue is
being worked on. -->

Ensure that the following pass:

- [x] `make format && make check` or via prek validation.
- [ ] `make test` passes locally
- [ ] `make test-e2e` passes locally
- [ ] `make test-ci-container` passes locally (recommended)
- [ ] GPU CI status check passes -- comment `/sync` on this PR to
trigger a run (auto-triggers on ready-for-review)

## Pre-Merge Checklist

<!-- These checks need to be completed before a PR is merged, -->
<!-- but as PRs often change significantly during review, -->
<!-- it's OK for them to be incomplete when review is first requested.
-->

- [ ] New or updated tests for any fix or new behavior
- [ ] Updated documentation for new features and behaviors, including
docstrings for API docs.

## Other Notes

<!-- Please add the issue number that should be closed when this PR is
merged. -->
- Closes #<issue>

---------

Signed-off-by: Sean Yang <seayang@nvidia.com>

Author: seayang-nv

docs/product-overview/data_synthesis.md

@@ -26,12 +26,9 @@ NeMo Safe Synthesizer adapts language models to understand and generate tabular
 - Generates new records that maintain statistical properties with no one-to-one mapping to original records
 - Supports various model sizes and architectures
 
-Two backends are available:
-
-| Backend | Description | When to use |
-|---------|-------------|-------------|
-| Unsloth | Optimized kernels for faster fine-tuning | Default -- use unless you need DP or a custom quantization setup |
-| HuggingFace | Standard PEFT training with 4-bit/8-bit quantization and optional differential privacy via [Opacus](https://opacus.ai/) | Required for differential privacy; also the fallback when Unsloth is unavailable |
+Two backends are available: Unsloth (default, faster) and HuggingFace
+(required for differential privacy). Both perform LoRA fine-tuning; see
+[Running -- Training](../user-guide/running.md#training) for a comparison.
 
 Three models have been extensively tested:
 

docs/product-overview/pipeline.md

@@ -46,12 +46,10 @@ Records are converted to a JSON format and tokenized for model training. The ass
 
 ### 4. Training
 
-The training stage fine-tunes a base LLM using LoRA (Low-Rank Adaptation). Two backends are available:
-
-| Backend | Description |
-|---------|-------------|
-| **HuggingFace** | Standard training with quantization (4-bit/8-bit), LoRA via PEFT, and optional differential privacy via [Opacus](https://opacus.ai/) |
-| **Unsloth** | Optimized training for faster fine-tuning |
+The training stage fine-tunes a base LLM using LoRA (Low-Rank Adaptation). Two
+backends are available -- Unsloth (default, faster) and HuggingFace (required
+for differential privacy). Both perform LoRA fine-tuning; see
+[Running -- Training](../user-guide/running.md#training) for details.
 
 Three models have been extensively tested:
 

docs/user-guide/getting-started.md

@@ -177,14 +177,9 @@ entity types, LLM classification setup, and SDK customization.
 ### 3. Training
 
 Fine-tunes a base LLM using LoRA (Low-Rank Adaptation). Two backends are
-available:
-
-| Backend | Description |
-|---------|-------------|
-| Unsloth | Optimized training for faster fine-tuning (auto-selected by default) |
-| HuggingFace | Standard training with quantization (4-bit/8-bit), LoRA via PEFT, and optional differential privacy via [Opacus](https://opacus.ai/) |
-
-If you enable differential privacy, the pipeline automatically switches to use the HuggingFace backend.
+available: Unsloth (default, faster) and HuggingFace (required for
+differential privacy). Both perform LoRA fine-tuning; see
+[Running -- Training](running.md#training) for a comparison.
 
 The default model is `HuggingFaceTB/SmolLM3-3B`. Safe Synthesizer has tested support for `HuggingFaceTB/SmolLM3-3B`, `TinyLlama/TinyLlama-1.1B-Chat-v1.0`, and `mistralai/Mistral-7B-Instruct-v0.3` (see [Configuration -- Training](configuration.md#training) for details on how to change the backend or model).
 

docs/user-guide/running.md

@@ -472,8 +472,10 @@ Two backends are available:
 
 | Backend | Description | When to use |
 |---------|-------------|-------------|
-| Unsloth | Optimized kernels for faster fine-tuning | Default -- use unless you need DP or a custom quantization setup |
-| HuggingFace | Standard PEFT training with 4-bit/8-bit quantization and optional differential privacy via [Opacus](https://opacus.ai/) | Required for differential privacy; also the fallback when Unsloth is unavailable |
+| Unsloth | LoRA fine-tuning with optimized kernels for faster training and lower VRAM usage. Uses Unsloth's `FastLanguageModel` for model loading and PEFT wrapping | Default -- use unless you need DP or a custom quantization setup |
+| HuggingFace | LoRA fine-tuning via PEFT with 4-bit/8-bit quantization support and optional differential privacy (DP-SGD) via [Opacus](https://opacus.ai/) | Required for differential privacy; also the fallback when Unsloth is unavailable |
+
+If you enable differential privacy, the pipeline automatically switches to the HuggingFace backend.
 
 Three models have been extensively tested: