Joakimpalm-Zen/Qwen3.8-27B-GSQ-RCO-scale-recovery-evidence

Qwen3.8-27B GSQ-RCO IQ3_S: scale-recovery evidence and code

<!-- BEGIN xyntetik-dataset-summary -->

Research evidence dataset. No model weights. Part of the collection Xyntetik Research: Model Surgery and Scale Recovery on this account, produced with Xyntetik Runner.

Dataset summary

Question tested. Whether retraining only the fp16 block scales of an existing IQ3_S file (integer codes untouched) by distillation against the BF16 parent brings it inside the house bar, and what it does on a public benchmark.

Models involved.

• —Produced artifact: https://huggingface.co/Joakimpalm-Zen/Qwen3.8-27B-GSQ-RCO-IQ3_S-recovered-GGUF (11,771,546,784 bytes, same byte length and layout as the source)
• —Source file: ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF, Qwen3.8-27B-GSQ-RCO-IQ3_S.gguf, sha256 64b53b64...
• —Parent: Qwen/Qwen3.8-27B revision 1d4bf0f2..., BF16

Method. Top-64 forward-KL distillation from the BF16 parent to the student's fp16 d/dmin fields (103,567,360 trainable scales in 399 tensors; embedding and output-head scales and all integer codes frozen), 300 steps, batch 4, lr 3e-4, seed recorded; then written back into a byte-identical layout and re-scored on 500 held-out prose positions and on HellaSwag (1,000 items) beside the unmodified source and the parent.

What each file contains.

• —RECOVERY.json: the provenance sidecar of the artifact (parent, source, exactly which fields changed, training record, verification)
• —evidence/run.json: the training run record (hyper-parameters, corpus hash, instrument hash)
• —evidence/train.jsonl: the per-step training log
• —evidence/training-corpus.manifest.json: the corpus manifest (what was trained on, by hash)
• —evidence/source-iq3s-vs-bf16-500.json and evidence/recovered-iq3s-vs-bf16-500.json: the 500-position fidelity harness records against the BF16 parent, before and after recovery
• —evidence/hellaswag-parent-1000.json, evidence/hellaswag-iq3s-source-1000.json, evidence/hellaswag-iq3s-recovered-1000.json: HellaSwag records for parent, source and recovered file
• —evidence/frontier-table.json: the public-quant frontier table the artifact is placed on
• —evidence/RESULTS.md: results of the parallel structural study (whether quantization and head surgery compose; depth and vocabulary levers) whose instruments this recovery reused
• —code/: the training, scoring and writing scripts (train_scales.py, write_scales.py, scaled_model.py, teacher_topk.py, benchmark_hellaswag.py, ...) with their unit tests
• —evidence/Qwen3.8-27B-GSQ-RCO-IQ3_S-recovered.recovery.json: the write-time verification record

Reproduction. the Reproduce section of the model card; the scripts in code/ with the hyper-parameters in evidence/run.json. Machine paths inside the records are the measurement box's and are left as recorded.

Result. Mean KLD 0.0450 and margin-qualified top-1 97.80% against the BF16 parent on the held-out positions, inside the house bar; the unmodified source file's matched row is in the same records so the reader sees exactly what recovery changed.

Links.

• —Model card: https://huggingface.co/Joakimpalm-Zen/Qwen3.8-27B-GSQ-RCO-IQ3_S-recovered-GGUF
• —Runner: https://github.com/Joakimpalm-Zen/xyntetik-runner
• —Source quantization method: GSQ (arXiv:2604.18556) and RCO (arXiv:2605.00649) <!-- END xyntetik-dataset-summary -->

These files are also present unchanged inside the model repository; this dataset is their first-class home so the evidence can be cited and browsed without a 12 GB download.