mkleinegger/tmp

Qwen3.8-27B GSQ + RCO — data artifact

Use this folder with qwen3.8-27b-gsq-rco-code.zip. Extract that ZIP next to this folder, or set GSQ_RCO_DATA_DIR to this folder's absolute path.

Contents:

• models/Qwen3.8-27B/: complete original checkpoint, tokenizer, model configuration, and model license; 18 safetensors shards, pinned to revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
• qwen3.8-27b-gguf-uniform/: all 12 uniform GGUF model files.
• calibration/fineweb_edu.pt: 1,024 token sequences of length 2,048, with sampling metadata in fineweb_edu.json.
• calibration/wikitext2/: saved train, validation, and test splits for offline evaluation.
• model-metadata/: Qwen3.8-27B SSM permutations and linear-module shapes.
• reference-assignments/: saved RCO assignments and the GSQ type mapping.
• model-manifest.json and input-manifest.json: checkpoint revision, available checkpoint hashes, and GGUF file-size records.
• data-copy-manifest.json: all 63 copied input files, sizes, and copy checks (full SHA256 for files up to 8 MiB; matching first/last 1 MiB for larger files).

Weights and GGUFs total about 170.4 GB, plus a small amount of dataset and metadata space. These are independent copies, not links into the original bundle. The derived candidate store is not included; build it with the code artifact's build_store.sh on a disk with about 591 GB additional capacity, plus room for later search caches and GSQ outputs.

The folder contains no Python environment or source repository. The original combined bundle is preserved.

Existing staged models in /home/mhelcig/data/Qwen3.8-27B-GSQ-RCO are excluded. In particular, the staged files named GSQ-RCO-IQ2_S and GSQ-RCO-IQ2_XS were checked and contain multiple quantization types across their linear layers: their filenames do not mean they are uniform models. Mixed-bit, block/code-pruned variants, and attic/ contents are not included.