NEFTune
Algorithms AIMech:NEFTune DRAFT
Aliases: Noisy Embedding Instruction Finetuning, NEFT Noisy Embedding Finetuning, Scaled Uniform Embedding Noise Instruction Finetuning
A regularization algorithm for language model instruction fine-tuning that perturbs input token embedding tensors X_emb in R^{B x L x d} during each training forward pass by adding i.i.d. uniform noise epsilon ~ Uniform(-1, 1) scaled by alpha / sqrt(L d), where L is sequence length, d is embedding dimension, and alpha > 0 is a tunable noise magnitude.
Identity Notes & Concept Scope
Distinct from adversarial embedding perturbation methods such as FreeLB (which require inner-loop gradient ascent steps to compute worst-case adversarial perturbations) and from parameter-space noise injection or dropout. NEFTune (Noisy Embedding Instruction Fine Tuning; Neel Jain, Ping-yeh Chiang, Yuxin Wen, John Kirchenbauer, Hong-Min Chu, Gowthami Somepalli, Brian R. Bartoldson, Bhavya Kailkhura, Avi Schwarzschild, Aniruddha Saha, Micah Goldblum, Jonas Geiping, & Tom Goldstein, ICLR 2024, arXiv:2310.05914) samples a single zero-overhead random vector epsilon ~ Uniform(-1, 1)^{B x L x d} per forward pass and computes X'_emb = X_emb + (alpha / sqrt(L d)) epsilon (Algorithm 1). Because E[epsilon_{t,j}^2] = 1/3, the perturbation has expected Euclidean norm alpha / sqrt(3) independent of sequence length L and embedding dimension d, reducing overfitting to the exact phrasing and length formatting of small instruction datasets without degrading factual knowledge or token-level nearest-neighbor semantics.
Claims & Grounded Evidence
4 Claims
Mechanism
#neftune-length-and-dimension-invariant-uniform-embedding-noise-scaling
For a token sequence of length L and embedding dimension d, sampling epsilon ~ Uniform(-1, 1)^{L x d} i.i.d. and forming X'_emb = X_emb + (alpha / sqrt(L d)) epsilon produces a non-adversarial perturbation with expected Euclidean norm alpha / sqrt(3) independent of L and d and zero additional backward passes.
Mechanism
#neftune-instruction-overfitting-mitigation-and-intra-token-preservation
NEFTune increases training loss on the instruction dataset while lowering held-out test loss and training-prompt ROUGE-L/BLEU memorization scores (preventing verbatim lock-in to training responses), while maintaining <= 0.4% nearest-neighbor vocabulary token flips and preserving n-gram log-diversity.
Application
#neftune-application-llama2-opt-qlora-instruction-tuning
Applied to LLaMA-2-7B across Alpaca, Evol-Instruct, ShareGPT, and OpenPlatypus, NEFTune boosts GPT-4 AlpacaEval win rate by +15.1 percentage points on average (including 29.79% -> 64.69% on Alpaca), improves RLHF-tuned LLaMA-2-Chat-7B (+7.30%) and LLaMA-2-70B (+13.78%), and improves QLoRA fine-tuning while preserving OpenLLM Leaderboard accuracy on MMLU, ARC, HellaSwag, and TruthfulQA.
Limitation
#neftune-limitation-verbosity-shift-and-safety-refusal-tradeoff
NEFTune induces substantially longer, more verbose completions (which correlate with LLM-judge preferences on AlpacaEval), yields smaller gains under 4-bit QLoRA than under full-parameter fine-tuning, and can affect safety-aligned checkpoints' refusal behavior against toxic outputs if applied without safety re-alignment.
⚠️ Curated Limitations & Negative Results
- Significantly increases output response length and verbosity, which can confound length-sensitive automatic evaluators such as GPT-4 AlpacaEval.
- Further fine-tuning RLHF-aligned chat models with NEFTune on instruction datasets without safety data can degrade refusal behavior against toxic prompts.
- Provides milder improvements under 4-bit QLoRA quantization than under full-parameter fine-tuning.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:ArtificialNeuralNetwork
Artificial Neural Network
|
related_match | NEFTune regularizes transformer neural network instruction fine-tuning by perturbing the input token embedding layer with scaled uniform noise. |
Evidence Sources & Primary Literature
3 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
r673-primary-pdf
|
Neel Jain, Ping-yeh Chiang, Yuxin Wen, John Kirchenbauer, Hong-Min Chu, Gowthami Somepalli, Brian R. Bartoldson, Bhavya Kailkhura, Avi Schwarzschild, Aniruddha Saha, Micah Goldblum, Jonas Geiping, Tom Goldstein
|
primary paper
full_text
|
ID: https://dblp.org/search/publ/api?q=NEFTune+Noisy+Embeddings+Improve+Instruction+Finetuning&format=xml
sha256:c5ac68e99a68…
|
r673-arxiv-abs
|
Neel Jain, Ping-yeh Chiang, Yuxin Wen, John Kirchenbauer, Hong-Min Chu, Gowthami Somepalli, Brian R. Bartoldson, Bhavya Kailkhura, Avi Schwarzschild, Aniruddha Saha, Micah Goldblum, Jonas Geiping, Tom Goldstein
|
primary paper
abstract_only
|
ID: https://arxiv.org/abs/2310.05914
sha256:5bbbf8fee65b…
|
r673-metadata-cr
|
Neel Jain, Ping-yeh Chiang, Yuxin Wen, John Kirchenbauer, Hong-Min Chu, Gowthami Somepalli, Brian R. Bartoldson, Bhavya Kailkhura, Avi Schwarzschild, Aniruddha Saha, Micah Goldblum, Jonas Geiping, Tom Goldstein
|
primary paper
abstract_only
|
ID: https://dblp.org/search/publ/api?q=NEFTune+Noisy+Embeddings+Improve+Instruction+Finetuning&format=xml
sha256:a478cfd4efac…
|
Ontology & Grounding Topology
Open in interactive network graph →Curation Provenance
Curator: Antigravity
Method: ai_assisted
Model: Gemini
Created: 2026-10-19T12:36:00Z
Updated: 2026-10-19T12:36:00Z
Source File:
data/curated/NEFTune.yaml
· Schema:
curated_concept.yaml
Explore across X-Mech
Related computational & biological Mechs:
Embedding Map
Network Graph
ProteinTraitsMech (ESM-2 pLM)
DUFMech (Domain pLM)
TraitMech
CultureMech
All 14 X-Mech Repositories