training-dynamics 开源项目推荐（50 个）

google /neural-tangents

Python中快速简单的无限神经网络

CVL-UESTC /FR-INR

CVPR 2024-Improved Implicit Neural Representation with Fourier Reparameterized Training

CVL-UESTC /IGA-INR

ICML2025-Inductive Gradient Adjustment for Spectral Bias in Implicit Neural Representations

MetaCircleAI /ComfyResearch

A platform for structuring AI research and studying physics of AI

jjbrophy47 /instance_based_interpretability

Existing literature about training-data analysis.

mainlp /explaind

A unified framework for attributing model behavior to model components, training data, and training dynamics.

pulkitgopalani /tf-loss-plateau

Code for "What Happens During the Loss Plateau? Understanding Abrupt Learning in Transformers" (NeurIPS 2025)

Shwai-He /Transformer-Geometry

[EMNLP 2026 Findings] Decomposing Transformer updates into parallel & perpendicular subspaces for geometric probing, compression, and training dynamics.

liuyz0 /FOCUS

Official repository for "FOCUS: First Order Concentrated Updating Scheme"

marlo-z /reversal_curse_analysis

Code for 'Towards a Theoretical Understanding of the 'Reversal Curse' via Training Dynamics'

braingpt-lovelab /backwards

Source code for <Probability Consistency in Large Language Models: Theoretical Foundations Meet Empirical Discrepancies>

dgcnz /relaxed-equivariance-dynamics

Code for "Effect of equivariance on training dynamics"

jaztulsi /optimizer-autopsy

Rewind a training run to the instant its loss spiked, localize the poisoned optimizer state, repair just that slice, and prove it via bit-identical replay forks. Instrument built + CUDA-verified on Kaggle T4; restarting on AMD MI300X. PyTorch research.

cosmaadrian /nli-stress-test

Official repository for the EMNLP 2024 paper "How Hard is this Test Set? NLI Characterization by Exploiting Training Dynamics"

tiexinding /NPM-Weibull-public

Transformer weight statistics & analysis: from pooled Weibull (k, λ) fits to row/column scale fields. Code, data, library.

flavio-gualtieri /eigengrokking

Do Hessian curvature spectra predict when a transformer will grok? 240 modular-arithmetic runs, stochastic Lanczos spectra and forecasts of grokking time from early training.

RayoHQ /attention-binding-a11y

TMLR 2026 | Mechanistic interpretability: attention-head binding (EB*) as a marker of concept emergence. 7 models, 5 architectures (Pythia 160M–2.8B, OLMo-1B, CRFM GPT-2, SmolLM3-3B, Qwen2.5-1.5B), 41 terms.

kaminoguo /desink

Effective rank, RankMe, E1, CKA and anisotropy on transformer hidden states are determined by one direction. The exact identity, and the attention sink behind it.

tiexinding /NPM-K-public

Cross-Family Convergence of Neural Network Weight Skeletons. Companion to Zenodo paper (10.5281/zenodo.19652706).

Tejas7007 /EMNLP_Findings_2026_Copying_Before_Suppression

Code, results, and Pythia-160M training checkpoints for "Copying Before Suppression: What Drives a Below-Chance Dip During Language Model Training?" (Findings of EMNLP 2026).