Private Llama35 serving adapters

35 original LoRA adapters and35 frozen Fleet spectral exports used by PROBE. Each of original/a00 through original/a34 and fleet/a00 through fleet/a34 is a PEFT adapter directory, not a merged base model.

metadata/adapters.json gives exact upstream repositories, resolved revisions, tasks, scaling configuration and SHA256 hashes. metadata/module-ranks.json gives the effective retained module ranks. Keep the original config scaling; compressed exports already fold original scaling into their factors (alpha/r=1). Global allocation uses weight spectra, not task labels or validation scores.

Effective compressed A+B parameters:398,457,856 (7.6613% of5,200,936,960). Padded checkpoint tensor parameters:922,746,880 (17.7419%). Padding must not be described as realized inference-compute or storage savings. Compression used the custom FlashTSQR GPU path; proof and frozen allocation are included.

Adapter aliases:14 r-three tasks,5 other classification adapters,10 multilingual NLI adapters and6 translation adapters into English. The base model, test examples, reference labels, credentials and environment binaries are excluded. These files alone do not establish task-quality equivalence or serving speedup.

Download for serving

from huggingface_hub import snapshot_download
snapshot_download("LoRAFleet/llama35-serving-adapters", local_dir="/workspace/llama35-adapters")

Authenticate using a read token with access to this private repository. Point your serving loader at the individual adapter subdirectories. The root of this multi-adapter archive is not itself a PEFT adapter. Native loaders may pad ranks; the custom compact execution path also needs effective ranks from metadata/module-ranks.json. Obtain the pinned base model separately under your existing access rights.

Upstream licenses

Original adapters retain their upstream provenance and applicable licenses; this private archive does not grant additional redistribution rights. Upstream model cards/license files are preserved alongside adapters where available. Consult the exact repositories and revisions in metadata/adapters.json and the Llama3.1 base model terms. No blanket replacement license is asserted.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for LoRAFleet/llama35-serving-adapters

Adapter
(3,057)
this model