Datasetpublic

neurips-ed2026-anon-checkpoints/checkpoint-zoo-metadata

LowRankArena Checkpoint Zoo Metadata Catalog This repository is a companion metadata catalog for the checkpoint zoo hosted in the linked Hugging Face Model repository. Each artifact-catalog record points to a checkpoint or reconstruction artifact at a fixed source commit. This repository does not duplicate or modify the referenced artifacts, and its CC BY 4.0 license does not apply to checkpoint files or upstream model weights. Source commit and coverage The full… See the full description on the dataset page: https://huggingface.co/datasets/neurips-ed2026-anon-checkpoints/checkpoint-zoo-metadata.

sourceHugging Facecc-by-4.0updated 2mo agoView on Hugging Face

0likes20downloads

Dataset Card

LowRankArena Checkpoint Zoo Metadata Catalog

This repository is a companion metadata catalog for the checkpoint zoo hosted in the linked Hugging Face Model repository.

Each artifact-catalog record points to a checkpoint or reconstruction artifact at a fixed source commit. This repository does not duplicate or modify the referenced artifacts, and its CC BY 4.0 license does not apply to checkpoint files or upstream model weights.

Source commit and coverage

The full catalog was constructed from eight family-level MANIFEST.csv files at the pinned source commit `d92dbff566e5de732959dce8cb8aaf3b12e8dfbd`.

It contains 300 manifest-listed artifact paths across six model families and 16 method labels. Of these, 289 are derived-model checkpoint paths and 11 are BlockPruner reconstruction-artifact paths. The reviewer-subset catalog contains 38 checkpoint paths, all from the Llama-3.1-8B family.

The catalog primarily documents released low-rank compression checkpoints. It also indexes structured-pruning artifacts: SliceGPT and LLMPruner entries point to derived-model checkpoints, while BlockPruner entries contain deletion-order metadata rather than standalone model weights.

Inventory

Model family	Catalog entries	Base-model identifier used for license mapping	Applicable upstream license
Llama-3.1-8B	69	meta-llama/Llama-3.1-8B	Llama 3.1 Community License Agreement
Llama-3.1-8B-Instruct	40	meta-llama/Llama-3.1-8B-Instruct	Llama 3.1 Community License Agreement
Llama-7B	71	huggyllama/llama-7b	LLaMA License Agreement (non-commercial research terms)
Llama2-7B	40	meta-llama/Llama-2-7b-hf	Llama 2 Community License Agreement
Qwen3-8B	40	Qwen/Qwen3-8B	Apache License 2.0
Qwen3-8B-Base	40	Qwen/Qwen3-8B-Base	Apache License 2.0

Files

—checkpoint_catalog.csv: 300 full-zoo artifact entries with explicit base-model, license, Responsible AI profile, and datasheet links.
—reviewer_subset_catalog.csv: entries listed in the root reviewer-subset manifest, with the same explicit metadata links.
—LICENSE_MATRIX.csv: base-model license scope and inheritance.
—RAI_DATASHEET.md: human-readable Responsible AI documentation.
—croissant_metadata.json: machine-readable core and RAI metadata.
—inventory_summary.json: inventory counts and validation scope.

Recovery and variant fields

checkpoint_variant preserves labels explicitly encoded in public artifact paths, including whitening_only, update_u256, whitening_then_update_u256, alpaca_recover, and prune_only. recovery_setting is alpaca_recover only for Alpaca recovery and none for all other entries. Labels such as update, update_c4, and whitening_then_update_u256 identify closed-form update variants, not recovery procedures.

Responsible AI

RAI_DATASHEET.md documents intended uses, unsupported uses, limitations, inherited base-model risks, and possible behavioral effects of compression and pruning. Catalog presence is not a safety, privacy, fairness, robustness, or deployment certification.

License scope

The CC BY 4.0 license applies only to the newly authored metadata and documentation in this catalog repository. It does not apply to checkpoint files, tokenizer files, configurations, code, benchmark data, or evaluation outputs. Use of checkpoint files remains subject to any applicable upstream model license and use terms listed in LICENSE_MATRIX.csv.

The catalog tables repeat the applicable base-model identifier and license information for record-level readability. LICENSE_MATRIX.csv, keyed by model_family, is the canonical family-level mapping.

rai_profile=shared_checkpoint_zoo_rai_profile identifies the shared Responsible AI profile documented at rai_datasheet_url; it does not represent checkpoint-specific risk evaluation.