HELIX-200M-SFT-ACTIVE

A short, low-LR chat-oriented SFT pass on top of the DPO-100 checkpoint will improve conversational behavior without disturbing the identity/voice DPO already bound in.

Run helix-200m-chat-tonight-v1: helix_200m preset, d_model 1024, 15 layers, 4 experts (top-2 routing), eag_r_max 256. Cosine LR 8e-06 → 8e-07, 61-step warmup, batch 2×8 at seq 1024, 2 epochs, 1,532 max steps.

• ◆Post-DPO SFT corpus and schedule (first round on top of DPO-100)

• →Live snapshot at step 1,270/1,532 (82.9%, epoch 2/2, elapsed 2:03:51): loss 1.4642, train PPL 4.32, held-out val PPL 4.43.
• →Throughput 2,677 tok/s; LR at snapshot 1.353e-06 (past peak, decaying per schedule); ETA to completion ~25 minutes.

• ·Val PPL (4.43) sits close to train PPL (4.32) with no divergence — a healthy fine-tune, not overfitting to the SFT set, at least at this snapshot.
• ·These numbers are not comparable to the base-model pretraining PPLs elsewhere on this site (27.50 held-out, etc.) — this is a short high-LR-decayed SFT pass on a tiny step budget, an entirely different regime.
• ·Observed directly from the operator's live training dashboard rather than from a saved log file — a genuine snapshot, not a registry artifact.

In progress, tracking healthy. Update this entry with final numbers once the run completes and gets evaluated on the frozen behavior suite.

Run the post-training bakeoff against this checkpoint once complete; decide whether it or DPO-100 (or the still-unadjudicated creator-unlikelihood-v3) becomes the standing model.