Skip to content

Latest commit

 

History

History
54 lines (43 loc) · 2.89 KB

File metadata and controls

54 lines (43 loc) · 2.89 KB

Token Distillation initialization

The production 8B initialization is complete and verified. Apertus-8B has untied input and output embeddings, so the procedure deliberately does more than learn one shared table.

Machine-readable authority: configs/initialization/token_distillation_8b.json.

Procedure

  1. Start from swiss-ai/Apertus-8B-2509@3162c996... and the published, uncpt 148,480-row layer-11 TokenDistil-Init checkpoint. Do not start from a CPT-trained checkpoint.
  2. Preserve all base IDs and append the 512 polytonic IDs 148480..148991 in dependency-safe merge order.
  3. Initialize the new input and output rows from their merge parents, with norm calibration against the modern added rows.
  4. Train the new input rows with layer-11 Token Distillation: 25 snippets per token, one epoch, batch 8, LR 1e-4, bf16.
  5. Because embeddings are untied, calibrate the 512 output rows separately with next-token cross-entropy for 400 steps at sequence length 512, LR 2e-4, gradient clip 1.0 and seed 20260729.
  6. Require exact preservation of every pre-existing input row, output row and non-embedding tensor, finite/nonzero new rows, and a zero-drift HF → Megatron TP=2 → HF round trip.

The fixed-ID adapter and production builder are preserved in fffoivos/train-apertus-with-glossapi:

Frozen evidence

Production verification:
/capstor/scratch/cscs/fffoivos/models/greek-cpt25b-init/
20260731T124000Z-cpt25b-v1/production_init_verification.json
sha256 7b4adb065f401305064cda8c45cc1b2c1430366ef8fbf5ee060f154be5b48e26

Round-trip verification:
/capstor/scratch/cscs/fffoivos/models/greek-cpt25b-init-roundtrip/
20260731T124000Z-cpt25b-v1/work/verification.json
sha256 06ddc69f4787f470b85d4be55992f898b98d8ab4fd8ab34a89a4d3bb96eb7e33

Megatron TP=2 checkpoint:
/capstor/scratch/cscs/fffoivos/models/greek-cpt25b-init-roundtrip/
20260731T124000Z-cpt25b-v1/megatron_tp2_r17patched

These are file hashes. They must not be confused with semantic tensor-tree digests that may also appear inside verification reports.