bluemorpholimited/Morpho-72B-MoE-DPO

Morpho DPO — preference alignment artifacts

📄 Reports: Full release report (PDF, 63 pp.) · Model card · Verification report

Stage 5 of the Morpho pipeline (display name; SFT backbone: Morpho-72B-MoE-Checkpoints).

Contract

• Per-expert DPO, same 64 experts, same one-example-one-expert mapping as SFT.
• Same convergence bar: patience-gated training, best-only weights, DONE markers.
• Real preference data only — 29,983 chosen/rejected pairs (HH-RLHF, HelpSteer2, UltraFeedback, Nectar), routed per expert by the v5 taxonomy with a 200-pair floor.
• DPO adapters initialize from the converged SFT adapters — never from scratch.

Layout (populated as the stage runs)

dpo/raw/        # preference pairs per source (chosen/rejected JSONL)
dpo/tagged/     # per-expert routed pairs + expert_id/domain audit fields
dpo/tokenized/  # tokenized chosen/rejected batches (dpo_batch_*.npy)
dpo/adapters/   # per-expert DPO adapters (sft/expert_XX pattern)
manifest.json   # per-expert pair counts + DPO losses

Status

✅ Stage 5 COMPLETE — 31/31 experts DONE (resumed 4x; all adapters + DONE markers on Morpho-72B-MoE-DPO).

Expert	Domain	DPO best loss	Converged step	Weights
02	Banking & Lending	1.01e-06	673	✅ DONE + weights
04	Accounting & Auditing	7.83e-08	317	✅ DONE + weights
05	Taxation & Compliance	5.73e-04	979	✅ DONE + weights
06	Real Estate Finance	0.0176	400	✅ DONE + weights
09	Criminal Law	0.0176	191	✅ DONE + weights
10	Civil Law	5.33e-06	587	✅ DONE + weights
11	Corporate Law	0.00105	525	✅ DONE + weights
12	Intellectual Property	1.38e-06	1695	✅ DONE + weights
13	International Law	6.38e-09	759	✅ DONE + weights
14	Tax Law	2.17e-10	1050	✅ DONE + weights
15	Regulatory Compliance	8.82e-04	607	✅ DONE + weights
17	Pharmacology	7.50e-16	739	✅ DONE + weights
18	Diagnostics & Imaging	3.23e-11	348	✅ DONE + weights
20	Neurology	0.00406	350	✅ DONE + weights
21	Cardiology	9.27e-05	660	✅ DONE + weights
22	Oncology	1.84e-08	1240	✅ DONE + weights
23	Public Health	3.15e-11	660	✅ DONE + weights
25	Systems & Infrastructure	3.61e-09	581	✅ DONE + weights
27	Databases & SQL	1.36e-05	1066	✅ DONE + weights
34	Biology & Life Sciences	2.37e-03	546	✅ DONE + weights
35	Mathematics	4.11e-05	1191	✅ DONE + weights
38	Civil Engineering	4.84e-06	194	✅ DONE + weights
40	Strategic Management	9.45e-07	577	✅ DONE + weights
43	Operations & Logistics	6.85e-04	327	✅ DONE + weights
47	Consulting	4.60e-05	772	✅ DONE + weights
56	Psychology	6.52e-03	388	✅ DONE + weights
57	Psychology (advanced)	6.05e-04	1050	✅ DONE + weights
59	Economics	3.77e-05	562	✅ DONE + weights
60	Philosophy	4.02e-03	943	✅ DONE + weights
62	Political Science	2.74e-02	183	✅ DONE + weights
63	Critical Analysis & Reasoning	3.94e-08	248	✅ DONE + weights

All 31 eligible experts are trained and published; 33 thin experts (<200 pairs) keep SFT-only per policy (documented above). Each row is verified from the expert's best_loss.txt + DONE.txt on this repo.

Where these adapters went

Stage	Artifact
6 · Merge	Morpho-72B-MoE-Merged — 64 adapters (DPO where available, SFT otherwise) merged with a gated TIES-style pass; PPL 7.69 → 7.83 (×1.02)
7 · Verify	GSM8K 76.7% · PPL 7.68 · 19.5 tok/s
8 · GGUF	Morpho-72B-MoE-GGUF — f16, Q8_0, Q5_K_M, Q4_K_M
9 · Distill	Morpho-7B-Distill · Morpho-1.5B-Distill

The 31 DPO adapters are also usable directly: load the SFT adapter from the checkpoint vault and the DPO adapter on top, per expert. Full pipeline narrative on the model card.

---

Release map

Stage	Artifact
Pipeline, gate history, benchmarks	Morpho-72B-MoE
4 · 64 converged SFT adapters	Checkpoints
5 · 31 DPO-aligned adapters	DPO
6 · Merged 72B weights	Merged
8 · Quantized builds	GGUF
9 · Distilled students	7B · 1.5B
Data and pipeline code	Morpho-72B-MoE-Data