À suivreGuide suivant
YaRN et extension de longueur de contexte
Technique
GUIDE IA du langage
Position Interpolation compresses the positional coordinates used by a pretrained RoPE-based model when extending its context.
Écoutez ce guideNarré par l'IA · ~2 min
The original study adapts models to longer windows with additional training, without establishing universal accuracy or eliminating long-sequence costs.
RoPE applies position-dependent rotations to query and key dimension pairs. Simply using much longer sequences can expose the model to positional and attention conditions beyond those learned during training. Basic Position Interpolation rescales m to m divided by s, where s is the ratio of new context length to the original length. This compresses the coordinate span rather than continuing unmodified positions into a much longer range. The model still sees a changed sequence and denser position spacing, so adaptation and evaluation matter. The original paper reports successful extensions in its experiments with additional fine-tuning. Uniform compression affects all frequencies together; methods such as YaRN treat frequency bands differently to address local-detail and range tradeoffs.
Evaluating a pretrained RoPE model at a longer target length after positional rescaling and adaptation.
Comparing interpolation with unmodified positional extrapolation in the same model.
Studying frequency-aware methods such as YaRN against uniform position compression.
Measuring task performance and resource use across the extended window.
Context-extension research compares positional schemes, training budgets, and long-context evaluations. A nominally supported length is not proof of consistent retrieval or reasoning quality throughout the window. Assess the intended tasks, model adaptation, memory, and latency rather than treating rescaling as free long-context capability.
Newsletter gratuite
Trois histoires d'IA vérifiées chaque matin de la semaine, écrites dans un anglais simple. Gratuit pour toujours, sans publicité.
Un e-mail chaque jour de la semaine. Désabonnez-vous en un clic. Nous ne vendons ni ne partageons jamais votre adresse.
Testez-vous
Commentaires instantanés sur chaque réponse et un certificat partageable avec une pièce d'identité vérifiable une fois que vous avez réussi un cours.
Soutenez l’enseignement gratuit de l’IA. AI Understanding est une organisation à but non lucratif 501(c)(3) : pas de publicité, pas de paywall, jamais. Faire un don
Position Interpolation compresses the positional coordinates used by a pretrained RoPE-based model when extending its context. The original study adapts models to longer windows with additional training, without establishing universal accuracy or eliminating long-sequence costs.
PI divise les indices de position par le facteur d'extension, mappant ainsi la séquence la plus longue dans la plage de distribution que le modèle a déjà apprise.
The rescaled coordinates enter RoPE's rotation-angle calculation.
Out-of-training positional behavior can affect attention; observed effects depend on the model and extension.
PI mappe m en m divisé par le facteur s = L'/L, compressant la plage la plus longue dans l'intervalle d'entraînement d'origine.
YaRN preserves high-frequency local dimensions and interpolates lower-frequency dimensions with a transition between bands.
Continuez à apprendre
Plus de guides sélectionnés pour ce sujet
À suivreGuide suivant
YaRN et extension de longueur de contexte
Technique