LFM2.5-2.6B-Fable-Finetune-Donor-Banks

Autonoma Fable donor expert banks

This repository contains four frozen BF16 expert banks selected for a controlled router-only research campaign on an LFM2.5-2.6B Fable host. It is not a standalone model: it has no host weights, tokenizer, or inference configuration. A small, provenance-bound Fuse router warm-start is included only as an initialization input.

All artifacts are non-routing. Their presence on the Hub does not authorize training, deployment, or promotion in Autonoma.

Artifacts

Total weight payload: 24,160,622,472 bytes (22.5013 GiB).

Each bank contains 32 selected experts for each of host layers 0–29: 960 expert instances and 2,880 canonical projection tensors. Qwen fused tensors were sliced along their expert axis and split into canonical gate_proj, up_proj, and down_proj records. No unselected expert payload was copied.

router-warmstart.safetensors contains 30 32 x 2048 BF16 gate matrices (3,935,280 bytes, SHA-256 99932db7a472efbdfe98678d3e8cb8f35369e84083678fdb481c53e523ff9b0c) extracted from Akahsizrr/fuse-1-Lite@d74e868eb8a6ec47b803a2e6f8ba451024f06d43. It contains no host or expert weights. router-warmstart.json binds the exact 11.44 GB source artifact hash and all source tensor names/shapes. Expert scales are deliberately reset by the training contract rather than copied from this artifact.

Provenance

The immutable selection is fable-donor-selected-banks.v1.json, SHA-256:

158689cca0679423925cfe3c2eefbd6a248045ed01d2f7dad4d74e05e3ad0d81

Every banks/<id>/bank.json binds its source revision, selection hash, byte count, and final artifact SHA-256. At Hub commit b2a9cf3c3592a10f409d9232496e8e1488f8b68f, all four declarations were independently matched against Hugging Face LFS metadata.

The extraction evidence originally recorded a private destination. The owner subsequently made the repository public. PUBLIC_VISIBILITY_ADDENDUM.json records that transition without rewriting the immutable extraction records.

Intended experiment

The banks are inputs to matched router-only transplants. Host and expert weights remain frozen; only router gates and bounded expert scales may train. The controlling design uses:

• frozen-host KL preservation;
• detached multi-scale expert probes;
• benefit-weighted routing targets with host-only as an explicit off class;
• temporary expert-eligible exploration that anneals away;
• conditional load balancing only after positive benefit is observed;
• stop-boundary, anti-loop, duplicate-tool, and succinctness objectives.

These files do not contain a trained hybrid and should not be presented as one.

Licensing and attribution

KAT-Coder-V2.5-Dev and Qwen3.5/3.6 declare Apache-2.0. Ornith-1.0-35B declares MIT. See THIRD_PARTY_NOTICES.md for source revisions, transformations, and attribution. The repository includes the Apache-2.0 license governing its Apache-derived portions; the permissive MIT obligations remain preserved in the notices.

No trademark rights are granted. Users remain responsible for reviewing the pinned upstream repositories and their terms.