Skip to content

Latest commit

 

History

History
81 lines (66 loc) · 4.27 KB

File metadata and controls

81 lines (66 loc) · 4.27 KB

Related work and project boundary

This comparison defines SFTGuard's scope; it is not an exhaustive novelty search. SFTGuard is a local correctness and regression gate around a fine-tuning pipeline. It does not replace trainers, parameter-efficient tuning methods, data-selection research, or general model benchmarks.

Training frameworks

TRL provides post-training algorithms and trainers, including supervised fine-tuning. Its official SFTTrainer documentation defines dataset formats, chat-template application, token-level loss, and assistant/completion masking. SFTGuard inspects evidence about those semantics; it does not implement or accelerate the training loop.

LLaMA-Factory and Axolotl are configurable fine-tuning systems spanning model loading, dataset preparation, distributed training, parameter-efficient methods, and evaluation integrations. SFTGuard is designed to sit beside such a system: it can audit exported local records and gate supplied base-versus-adapter results, but it does not launch their jobs or claim configuration parity with them.

Parameter-efficient fine-tuning

PEFT implements LoRA and other parameter-efficient methods. Its LoRA API supports target modules, per-module rank and alpha patterns, multiple initialization methods, quantization-aware workflows, and automatic target selection. SFTGuard does not propose a new adapter, choose an optimal rank, or allocate modules. It checks pipeline correctness and evaluates an explicit release contract for an adapter that another tool produced.

Prediction and experiment screening

TuneAhead predicts full fine-tuning performance using static dataset descriptors and features from a short probe run. SFTGuard deliberately makes a different claim: it does not estimate a future score. It verifies bounded, observable properties and returns ABSTAIN when the evidence required for a gate is missing.

Data selection and processing

LESS selects influential examples for targeted instruction tuning using optimizer-aware low-rank gradient features. DataInf estimates data influence for LoRA-tuned language and diffusion models. These methods rank training examples by estimated utility. SFTGuard does not rank examples or assert that a dataset is high quality; its data checks target deterministic conditions such as schema errors, exact duplication, conflicting answers, split leakage, and bounded privacy signals.

Data-Juicer is a broad data processing system with filtering, analysis, transformation, and distributed execution operators for foundation-model data. SFTGuard is narrower: it uses a small frozen audit contract and emits a release decision artifact. It is not a general cleaning or data-transformation pipeline.

Evaluation and forgetting

lm-evaluation-harness runs a large catalog of language-model evaluations. TRACE, introduced in the TRACE paper, benchmarks continual learning across multiple tasks and measures degradation in general and instruction-following abilities. SFTGuard does not reproduce those benchmark suites. It consumes paired evaluation records chosen by the user and applies explicit target and retention budgets; the resulting Regression Card is valid only for that supplied contract.

Practical distinction

SFTGuard's intended contribution is the integration of four conservative software checks: exact training-data semantics where they can be verified, deterministic fault findings, paired regression budgets, and shareable content-addressed provenance. Existing projects above can provide the trainer, adapter, data pipeline, selector, or benchmark. SFTGuard records whether a specific proposed release met its declared evidence gates, without converting missing evidence into success or claiming that the release is safe in general.