Module FTDD-01 — MiniCPM Family (OpenBMB)

MiniCPM Family (OpenBMB)

The course's on-ramp hero: a family of small, genuinely-open models (MiniCPM5-1B, 3-4B, V 4.6, o 4.5) from Tsinghua + ModelBest, with the Ultra* datasets, Apache-2.0, and first-party fine-tuning via SWIFT and LLaMA-Factory.

minutes

artifacts

sub-sections

Why a 1.08B-parameter model from OpenBMB is the base every early module loads: cheap iteration, auditable provenance, and Apache-2.0 with no license friction. The modality progression (text → vision → omni), the Ultra* datasets that make the data modules inspectable, and the SWIFT-vs-LLaMA-Factory decision — the full MiniCPM stack, examined as the course's default teaching vehicle.

Key Claims

Load-Bearing Claims

MiniCPM is the course's default base because it is small, open, and license-free. MiniCPM5-1B (1.08B, Intelligence Index 17.9) fine-tunes in minutes on a consumer GPU — demonstrable proof of the FT00 thesis (steering, not teaching). Apache-2.0 with no MAU or field-of-use clauses means students can ship without license friction.

The family spans the modality axis, not a capability axis. MiniCPM5-1B (text) → MiniCPM3-4B (denser text) → MiniCPM-V 4.6 (+ SigLip-400M vision encoder) → MiniCPM-o 4.5 (omni, full-duplex). Each step adds a modality or capability the previous lacked; SigLip-400M is the load-bearing component that turns a text base into a vision-language model.

The Ultra* datasets (UltraChat, UltraFeedback arXiv:2310.01377, Ultra-FineWeb) are the course's data reference, not just MiniCPM's model reference. They are open, documented, and reproducible — so the course can point at a concrete preference pair and say 'this is the signal DPO optimizes.' A closed dataset cannot match this as a teaching tool.

SWIFT vs LLaMA-Factory is decided by modality, not preference. Text models (5-1B, 3-4B) produce equivalent results in both. Vision (V) and omni-modal (o) favor SWIFT — ModelBest/OpenBMB's first-party framework — for native modality handling, first-party chat templates, and example scripts maintained by the people who built the model.

After This Module

Place the MiniCPM family on the open spectrum and explain why it is the course's default base (open-data + Apache-2.0 + consumer-hardware-sized).

Map the family — MiniCPM5-1B, MiniCPM3-4B, MiniCPM-V 4.6, MiniCPM-o 4.5 — to the modality gap each fills, and explain SigLip-400M's role as the vision encoder.

Trace the Ultra* datasets (UltraChat, UltraFeedback, Ultra-FineWeb) to the steering techniques they enable (SFT, DPO, CPT) and cite the arXiv provenance (2310.01377).

Defend, with the Nature Communications paper (s41467-025-61040-5), why an open-data family is the right teaching vehicle: cheap iteration, auditable provenance, no license friction.

Choose between SWIFT and LLaMA-Factory for fine-tuning a MiniCPM base and explain the modality-based decision rule.

Artifacts

Teaching Document

~2,400 words; the MiniCPM family (5-1B, 3-4B, V 4.6, o 4.5), the Ultra* datasets (UltraChat/UltraFeedback/Ultra-FineWeb), why MiniCPM is the teaching vehicle (Nature Comms s41467-025-61040-5), and the SWIFT vs LLaMA-Factory decision

READ

Diagrams

4 diagrams — the MiniCPM family (modality axis), the Ultra* datasets → steering layer map, why MiniCPM is the teaching vehicle (three properties), SWIFT vs LLaMA-Factory decision

READ

Slide Deck

9 slides — title, the family, MiniCPM5-1B specs, Ultra* datasets, teaching-vehicle properties, SWIFT vs LLaMA-Factory, anti-patterns, what you can now do

READ

Teaching Script