Qwen3.8-27B Joint ARC Research Checkpoint
Standalone BF16 checkpoint with the trained LoRA adapter merged into the base.
No separate PEFT adapter is required. Base revision:
1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
Main training completed11 optimizer updates with528 distinct selected records:
259 public official ARC2 training tasks,102 synthetic programs,102 repairs,
64 synthetic action records,and1 public ARC3 observed-transition record.
This is not the entire646-record available pool. Prior adapter warm-start and
current-hardware pilots are separately documented in the local run report.
No benchmark evaluation,ARC score or quality-improvement claim is made.
Complete contexts and BF16 precision were retained. All1184 HF model tensors
were validated against exact PEFT safe-merge arithmetic;496 were adapted.
Original auxiliary MTP tensors are preserved under auxiliary/ but unused by
the current HF inference architecture. Standalone HF reload was verified;
Kaggle/vLLM fit and throughput were not measured.
See artifact_manifest.json for file hashes and exact artifact identity.
No datasets,credentials,optimizer state or private game sources are included.
Upstream Apache-2.0 licensing applies.