deepseek-v4-mini-300M-sft-v1

Experimental / research artifact — not a usable model. First SFT leg on cpt-fineweb, re-acquiring the parent-family distribution (DeepSeek-3.1/R1-generated traces) after the broad-data detour. Local RTX 5090, 2026-07-14.

Run

• 800 dense steps (--phases 0), seq 4096, batch 1, grad-accum 8, grad-checkpointing, Muon, WSD lr 1e-4, ~26M tokens
• Loss 8.18 → 5.23
• W&B: rsprw6uc

Caveats

• Dense-only training on a phase-B checkpoint: the Lightning Indexer was not re-calibrated (no phase A/B after) — per the recovery kit, backbone drift without indexer KL risks CSA decalibration. Continuations should re-run phases A/B.
• Coherence smoke at loss 5.23: incoherent. The loss-vs-coherence reference point for this family is the H100 run (~3.8 on the same data without the fineweb detour).

Continued in sft-v2.