Shared algorithms, model-specific releases

September 21, 2026 ยท View on GitHub

sliders-conceptmod is the research home and owner of the installable concept-slider-core package. Small model repositories depend on an immutable core revision and publish their own model integration, recipe, evidence and samples. Hugging Face hosts their artifacts.

flowchart TB
  core[sliders-conceptmod: research and concept-slider-core]
  anima[anima-concept-sliders: Anima integration and release]
  hf[Hugging Face: checkpoints and samples]
  core -->|pinned Python dependency| anima
  anima --> hf
  future[Future verified model integrations]
  core -.-> future
AreaShared coreModel repository
Particle adapterSoft routing and bottleneck MLPProjection names, rank, particle cloud ownership, hooks, strength controls
LearningPaired losses, critic, VIC regularizer and noise functionFrozen targets, normalization policy, optimizer recipe, update loop and exact replay
Ordinary LoRA fittingDual ridge solveActivation capture, calibration budget, up matrix, export names, evaluation
EvidenceAlgorithm tests and extraction provenanceWeight hashes, prompts, samples, model license, runtime lock and integration tests

The initial consumer is Anima. Its native runtime, trainer, ComfyUI plugin and distiller use the shared package. The byte-identical reference extraction preserves Anima's checkpoint fingerprint: its compatibility shim delegates to the core and its provenance code hashes the actual core implementation file. The release retains the original trainer and model-specific recipe.

YuE2 and Music 3 are not migrated in this change. Their current releases remain reproducible through their existing source and pins. They illustrate why model collection and evaluation stay outside the package: diffusion velocities and autoregressive hidden states require different targets and different held-out tests.

Research and fitting

With particle cloud P and a projected input u, routing computes z = softmax(R(u) @ P.T / sqrt(particle_dim)) @ P. The branch predicts up(F(concat(u, z))). Particle training compares paired noisy real and fake errors through relativistic logistic losses; it regularizes the particle cloud and critic. The model adapter defines what those errors mean.

To fit an ordinary LoRA, Anima keeps the teacher's up matrix U and solves for the down matrix A from calibration inputs X and routed bottleneck outputs H:

lambda = 0.01 * max(mean(diag(X @ X.T)), 1e-8)
A = solve(X @ X.T + lambda * I, H).T @ X
delta(x) = (x @ A.T) @ U.T

This minimizes a regularized calibration objective. It does not establish the best visual or audio result. Anima fixes the ridge fraction before development evaluation; its smaller projection error is not an image-quality score. Its published distills show milder lighting changes than the original particles. The complete Anima formulation and distillation recipe define the current experiment. Existing YuE2 research remains in the root README.

Adding an algorithm or model

  1. Keep tensor-level algorithms independent of model imports and checkpoint keys.
  2. Name a new recipe or API when the objective, normalization or solver changes; do not silently change an established release's defaults.
  3. Test gradients and checkpoint-state compatibility, and compare outputs against the prior implementation before switching a consumer.
  4. Fit on training data. Use development data for declared selection, and keep final-test examples out of fitting and selection. Publish perceptual samples alongside internal metrics.
  5. Pin a full core commit in the consumer and record the algorithm file hashes. Publish weights and evidence under a new release if the behavior changes.

The package version identifies the API family; the Git revision and file hashes identify the exact research implementation. Existing tagged releases remain valid even as research continues here.