bluemorpholimited/Morpho-180B-Data

The fuel behind MORPHO-180B. 822,745 examples from 17 public sources, every one routed to exactly one of 64 domain experts. Raw text, domain tags, tokenized batches, and all the code — audit any byte, reproduce any step.

✅ All nine pipeline stages are complete. These bytes trained 64 converged SFT experts, then 31 DPO experts, then produced the merged 72.7B model, four GGUF builds, and two distilled students.

Refinery

flowchart LR
    A["1 - Raw<br/>17 sources, JSONL"] --> B["2 - Tagged<br/>domain tags + provenance"]
    B --> C["3 - v5 map<br/>1 example = 1 expert"]
    C --> D["4 - Tokenized<br/>82 NumPy batches"]
    D --> E["5 - 64 experts<br/>822,745 routed"]
    E --> F["SFT + DPO<br/>64 + 31 experts"]
    style C fill:#00f0ff,stroke:#39ff88,color:#000
    style F fill:#39ff88,stroke:#00f0ff,color:#000

Taxonomy v5 is canonical — the same expert IDs and names used by the model, the checkpoint vault, and every training run. Eight sectors × eight specialists: Finance, Law, Medicine, Technology, Science, Business, Arts, Humanities.

What's in here

Path	Files	Contents
raw/	37	17 source corpora (JSONL) + their tagged twins + 3 override rebuild sets
tagged/	20	per-source domain routing with domain_id / domain_name / source
tokenized/	85	82 batch_XXXX.npy tokenized shards + manifest
code/	13	the actual pipeline: taxonomy, build, SFT, DPO, merge, GGUF, verify, distill
manifest.json	1	total counts, per-expert counts, taxonomy snapshot

Shards use a 152,064-token vocabulary and a 2,048-token training length.

Sources

BoolQ · GSM8K · HellaSwag · HelpSteer2 · MATH-500 · OpenAssistant · OpenOrca · SQuAD · Sujet-Finance-177k · alpaca · dolly-15k · finance-alpaca · fineweb-edu · hh-rlhf · twitter-financial-sentiment · twitter-financial-topic · yahoo-finance-news

Three override sets (rebuild_31_embedded, rebuild_52_translation, rebuild_54_stories) top up thin experts with real domain text so no expert trains on a starvation diet.

Distribution

Statistic	Examples
Total routed	822,745
Smallest expert	606
Median expert	6,334
Largest expert	92,366
Experts with ≥ 10K examples	23

Counts are uneven by design: the corpus is what the web actually contains, not a synthetic balance. Thin experts are handled by policy, not by duplication — experts with fewer than 200 preference pairs stay SFT-only in Stage 5 (33 of 64).

Preference data (Stage 5)

29,983 chosen/rejected pairs (HH-RLHF, HelpSteer2, UltraFeedback, Nectar), routed one-pair-one-expert with the same v5 taxonomy and a 200-pair floor. Tokenized DPO batches and per-expert adapters are published in Morpho-180B-DPO; 31 experts cleared the floor and converged.

Load it

import json
from huggingface_hub import hf_hub_download
import numpy as np

repo = "bluemorpholimited/Morpho-180B-Data"

# one tagged shard
p = hf_hub_download(repo, "tagged/alpaca_tagged.jsonl", repo_type="dataset")
rows = [json.loads(line) for line in open(p, encoding="utf-8")]
print(rows[0]["domain_id"], rows[0]["domain_name"], rows[0]["source"])

# one tokenized batch
b = hf_hub_download(repo, "tokenized/batch_0000.npy", repo_type="dataset")
arr = np.load(b, allow_pickle=True)
print(arr.shape, arr.dtype)

Reproducibility

code/ is the same code that ran on the training boxes, including the TIES-style merge (merge_export.py) and the quality gate (verify_model.py). The merge gate history — plain mean rejected at 39.1× base perplexity, caps 2% / 1.1% rejected, 0.6% shipped at ×1.02 — is documented on the model card.

Provenance and licence

Corpora are public datasets, redistributed under their original licences; the tagging, routing, tokenization, and training code is Apache 2.0. If you use this data, cite the upstream sources as well as this repository.

EVERY BYTE ROUTED. EVERY STEP REPRODUCIBLE.

---

Release map

Stage	Artifact
Pipeline, gate history, benchmarks	Morpho-180B
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-180B-Data