Add Qielbas Tiny 9M M1 (Labqoat)
Adds Qielbas Tiny 9M M1 under Labqoat.
9,557,716 parameters; four GatedDeltaNet2 blocks and one causal attention block. Pretrained on 6.7B tokens using Muon for hidden matrices and AdamW for the remaining parameters. The release uses a late-checkpoint FP32 average followed by five epochs of ARC-Easy/Challenge training-split adaptation with AdamW and general-text replay. Adaptation epochs were selected using held-out validation; ARC test questions were excluded from adaptation.
| Benchmark | Result |
|---|---|
| BLiMP, 67000 pairs | 75.3223880597015% |
| ARC-Easy test, 2376 questions | 48.274410774410775% |
| WikiText-2 byte perplexity | 2.907906720918924 |
Self-reported results using the frozen Glint-1.3 protocol: 256-token contexts, no BOS, unnormalized likelihood. Reproduced Efficiency under retained normalization: 80.142581.
Thanks for the detailed submission. The PR provides the architecture, training setup, benchmark details, and linked reports needed to review Qielbas Tiny 9M M1. The separation of pretraining details and ARC adaptation methodology is useful context for evaluating the entry.