Ionel2023/runpod-workspace-archive-2026-09-07

Qwen2.5 sparsification experiment checkpoints

Research artifacts from experiments that prune Qwen2.5 weights while constraining the change in the model's output distribution. The collection preserves intermediate pruning stages, recovery states, teacher probability tensors, and source-model cache objects for reproducibility and analysis. Checkpoints are experimental intermediate results, not validated production models.

The completed archive contains 123 objects totaling 277,823,892,009 bytes (277.824 GB). Every object was checked against its original SHA-256 and byte count, and a checkpoint restoration was tested.

Contents and layout

• —objects/trimlab_sweep/: stage and cut checkpoints for 0.5B, 1.5B, 3B, 7B, and 14B runs.
• —objects/trimlab_recovery*/: recovery/resumption state.
• —objects/sparse_repair_lab/: sparse-repair experiment tensor state.
• —objects/trimlab_persistent/: teacher-output tensors, pruning runs, and pinned Qwen2.5-32B source-model cache blobs.
• —objects/hf_trimlab/: original Qwen2.5-7B/14B model cache blobs.
• —manifest.json: original relative paths, byte sizes, SHA-256 hashes, and immutable revisions for verified files. The archive is complete.

Stage and recovery files use PyTorch checkpoint serialization; teacher tensors and source weights use safetensors. The checkpoint collection supports inspection of sparsity trajectories, preservation of KL distributions, and comparisons between pruning and recovery stages. The archive itself does not establish model quality or reproducibility of any particular reported metric.

Original Qwen2.5 0.5B and 1.5B models remain local and are excluded from this archive. Source code, workspace configuration, credentials, and unrelated reports are not published by this archival task.

Original models and attribution

The original models are from Qwen: Qwen2.5-0.5B-Instruct, Qwen2.5-1.5B-Instruct, Qwen2.5-3B-Instruct, Qwen2.5-7B-Instruct, Qwen2.5-14B-Instruct, and Qwen2.5-32B-Instruct. Refer to each original model repository for its model card and applicable license. Source cache paths preserve original model and revision identifiers; a blob filename containing 64 hexadecimal characters is its original SHA-256 digest. Upstream license files accompany source weights in upstream-licenses/ where available. No additional license grant for unpublished experiment code is implied.

Retrieval

Choose the desired original path in manifest.json, then download its remote_path from the recorded immutable revision. Validate its byte count and SHA-256 before use. This avoids downloading the whole collection.

The included restore.py implements streaming, hash-checked restoration into /workspace using a local copy of manifest.json named state.json in /workspace/archive_ops/2026-09-07. It refuses to overwrite existing files. Public downloads do not require a Hugging Face token.

Only files successfully committed and verified against their local SHA-256 are recorded as verified. The manifest's complete flag indicates whether the entire planned archive has completed.