Course: Course 3 — LLM Fine-Tuning Masterclass
Deep-Dive: FTDD-01 — MiniCPM Family (OpenBMB)
Duration: ~30 minutes (spoken at ~140 wpm)
Format: Verbatim transcript with [SLIDE N] cues. Read aloud or use as speaker notes.
[SLIDE 1 — Title]
Welcome to deep-dive FTDD-01, the MiniCPM Family from OpenBMB. This is the first of three deep-dives that put real model families and tooling under the microscope. If you have done modules FT zero-zero and FT zero-two, you already know MiniCPM by reputation — it is the on-ramp hero, the base every early lab loads. This deep-dive tells you who made it, what is in the family, and why a one-billion-parameter model from a Tsinghua lab is the right place to learn fine-tuning.
[SLIDE 2 — The family]
OpenBMB stands for "Open Lab for Big Model Base." It is a collaboration between Tsinghua University and a company called ModelBest, and its product is the MiniCPM family — a roster of small models, all Apache-two-point-zero, all released with open training data.
The family is not one model. It is a modality axis. Start with MiniCPM five dash one B — a dense one-point-oh-eight-billion-parameter text model, the on-ramp. Add density and you get MiniCPM three dash four B, the mid-size text base. Add a vision encoder — specifically SigLip-four-hundred-M, a four-hundred-million-parameter image encoder — and you get MiniCPM-V four point six, the vision-language variant. Add audio and full-duplex streaming, and you get MiniCPM-o four point five, the omni-modal model.
Each step adds a modality or a capability the previous member lacked. SigLip-four-hundred-M is the load-bearing component that turns a text base into a vision-language model. And because SigLip is itself an open, well-understood encoder, the vision-language stack is auditable down to the image encoder, not just the language backbone.
[SLIDE 3 — MiniCPM5-1B]
The course's default is MiniCPM five dash one B. One-point-oh-eight billion parameters, released May twenty twenty-six, scoring seventeen point nine on the Intelligence Index — strong for a model that fits comfortably in two gigabytes of VRAM.
Why one billion and not seven billion? Because of the thesis from FT zero-zero. Fine-tuning steers behavior; it does not inject knowledge. And that thesis is demonstrable on a one-billion model in minutes. You do not need a frontier model to learn that SFT changes format, that DPO shifts preference, or that abliteration deletes a direction. You need a model cheap enough to run the experiment ten times — to break it, reset it, and break it again. That is what teaching fine-tuning requires.
This is the model the FT zero-zero lab loads, the one FT zero-eight's QLoRA walkthrough fine-tunes, and the one every VRAM calculation in FT zero-one uses as its worked example. It is not the smartest model in the family. It is the most teachable.
[SLIDE 4 — The Ultra* datasets]
The reason MiniCPM is the course's data reference, not just its model reference, is the Ultra-star family of open datasets. Three of them matter. UltraChat — large-scale multi-turn dialogue, the SFT dataset. UltraFeedback — preference and feedback pairs, the DPO dataset, introduced in the paper at arXiv twenty-three-ten point oh-one-three-seven-seven. And Ultra-FineWeb — a curated web pretraining mix, for continued pretraining or base distribution work.
Each dataset maps to a steering layer. UltraChat to Layer three, SFT. UltraFeedback to Layer three, DPO. Ultra-FineWeb to Layer one, the base. And each appears in a specific course module — UltraChat in FT zero-four and FT twelve, UltraFeedback in FT zero-five and FT thirteen, Ultra-FineWeb in FT zero-six.
The point is not that these are the best datasets. It is that they are open, documented, and reproducible. When FT thirteen teaches DPO, it teaches it on UltraFeedback because you can go read the dataset, find a concrete preference pair, and say: this is the signal DPO is optimizing. With a closed dataset, that pointer is impossible. Open data gives you a steering wheel you can inspect. That is why the data modules keep returning to Ultra-star.
[SLIDE 5 — Why MiniCPM is the teaching vehicle]
The Nature Communications paper — identifier s-four-one-four-six-seven dash oh-two-five dash six-one-oh-four-oh-five — makes the case formally. Three properties make MiniCPM the ideal teaching vehicle, and they map directly onto the course's needs.
First, cheap iteration. A one-billion model fine-tunes in minutes on a consumer GPU. A student can run the full SFT, DPO, quantize, deploy loop in a single afternoon. You cannot teach fine-tuning on a model that takes six hours per run.
Second, auditable provenance. Open weights, open data, open recipe. Every claim about what the model saw can be checked against the published corpus. This is the FT zero-two auditability predicate, made concrete on the base the course actually uses.
Third, Apache-two-point-zero with no friction. No field-of-use restrictions, no monthly-active-user clauses, no named-product carve-outs — contrast Llama's Community License. A student can build on MiniCPM, ship a product, and never consult a lawyer about the model license. For a course that wants students to ship things, this matters.
If you want a community companion to the formal paper, Sam Witteveen's walkthrough — at youtu dot be slash o-x-one-m-W-two-N-nine-Z-underscore-Y — covers the family overview and the fine-tuning workflow in a video format that complements this deep-dive.
[SLIDE 6 — SWIFT vs LLaMA-Factory]
Two frameworks dominate MiniCPM fine-tuning, and the course references both. The choice is a real engineering decision.
SWIFT — Scalable lightWeight Infrastructure for Fine-Tuning — is ModelBest and OpenBMB's own framework. The first-party option. Its advantage is native, first-day support for every MiniCPM variant, including the vision and omni-modal models. Its disadvantage is that it is MiniCPM-flavored. It supports other models, but its center of gravity is the OpenBMB ecosystem.
LLaMA-Factory is the general-purpose framework — a unified interface for fine-tuning dozens of model families. Its advantage is breadth. If your team already uses it for Llama, Qwen, and Mistral, adding MiniCPM is a config change. Its disadvantage is that support for the newest MiniCPM variants — especially the omni-modal models — can lag SWIFT by weeks to months.
Here is the decision rule. If you are working with MiniCPM-V or MiniCPM-o, or you want first-party assurance that the chat template and modality handling are correct, use SWIFT. If you are fine-tuning MiniCPM five dash one B or three dash four B as part of a multi-model pipeline, use LLaMA-Factory. Both produce equivalent results on the text models. The difference is ecosystem fit.
Neither is the framework this course uses for its core training-loop deep-dives — FTDD-zero-four covers TRL, FTDD-zero-five covers Axolotl, FTDD-zero-three covers Unsloth. SWIFT and LLaMA-Factory are the MiniCPM-specific entry points.
[SLIDE 7 — Anti-patterns]
Three anti-patterns to leave with.
First, treating MiniCPM five dash one B as a production frontier model. It is not. It is a small, capable, open model — ideal for teaching, prototyping, and edge deployment. It will lose head-to-head capability comparisons against seventy-billion-plus models. The course uses it because it is cheap to iterate on, not because it is the strongest model available. Match the base to the deployment — that is FT zero-three.
Second, assuming the Ultra-star datasets are clean because they are open. Open does not mean clean. UltraChat, UltraFeedback, and Ultra-FineWeb are open and documented, which means you can audit them — not that the audit has already been done. Run the same PII sweep, deduplication, and decontamination you would run on any dataset. Openness gives you the ability to vet. It does not do the vetting.
Third, picking SWIFT or LLaMA-Factory without considering the modality. The text models fine-tune equivalently in both. The vision and omni-modal models do not. Choosing LLaMA-Factory for an omni-modal fine-tune because it is what you already use is a reasonable instinct that will cost you days of template-debugging.
[SLIDE 8 — What you can now do]
You can now place MiniCPM on the open spectrum and explain why it is the course's default base. You can map the family — five dash one B, three dash four B, V four point six, o four point five — to the modality gap each fills, and explain SigLip-four-hundred-M's role. You can trace the Ultra-star datasets to the steering layers they enable and cite the arXiv provenance. You can defend, with the Nature Communications paper, why an open-data family is the right teaching vehicle. And you can choose between SWIFT and LLaMA-Factory and defend the tradeoff.
The lab asks you to load MiniCPM five dash one B, apply a format-steering LoRA via SWIFT, and measure the before-and-after behavior shift. The steering thesis, felt on the course's default base, in under thirty minutes.
Next, deep-dive FTDD-zero-two: OLMo two and three plus Tulu three, from the Allen Institute. The fully-open comparison — the research-oriented counterpoint to MiniCPM's product-and-edge orientation.
End of deep-dive FTDD-01. Duration: approximately thirty minutes at one-hundred-forty words per minute.
# Teaching Script — Deep-Dive FTDD-01: MiniCPM Family (OpenBMB) **Course**: Course 3 — LLM Fine-Tuning Masterclass **Deep-Dive**: FTDD-01 — MiniCPM Family (OpenBMB) **Duration**: ~30 minutes (spoken at ~140 wpm) **Format**: Verbatim transcript with `[SLIDE N]` cues. Read aloud or use as speaker notes. --- [SLIDE 1 — Title] Welcome to deep-dive FTDD-01, the MiniCPM Family from OpenBMB. This is the first of three deep-dives that put real model families and tooling under the microscope. If you have done modules FT zero-zero and FT zero-two, you already know MiniCPM by reputation — it is the on-ramp hero, the base every early lab loads. This deep-dive tells you who made it, what is in the family, and why a one-billion-parameter model from a Tsinghua lab is the right place to learn fine-tuning. [SLIDE 2 — The family] OpenBMB stands for "Open Lab for Big Model Base." It is a collaboration between Tsinghua University and a company called ModelBest, and its product is the MiniCPM family — a roster of small models, all Apache-two-point-zero, all released with open training data. The family is not one model. It is a modality axis. Start with MiniCPM five dash one B — a dense one-point-oh-eight-billion-parameter text model, the on-ramp. Add density and you get MiniCPM three dash four B, the mid-size text base. Add a vision encoder — specifically SigLip-four-hundred-M, a four-hundred-million-parameter image encoder — and you get MiniCPM-V four point six, the vision-language variant. Add audio and full-duplex streaming, and you get MiniCPM-o four point five, the omni-modal model. Each step adds a modality or a capability the previous member lacked. SigLip-four-hundred-M is the load-bearing component that turns a text base into a vision-language model. And because SigLip is itself an open, well-understood encoder, the vision-language stack is auditable down to the image encoder, not just the language backbone. [SLIDE 3 — MiniCPM5-1B] The course's default is MiniCPM five dash one B. One-point-oh-eight billion parameters, released May twenty twenty-six, scoring seventeen point nine on the Intelligence Index — strong for a model that fits comfortably in two gigabytes of VRAM. Why one billion and not seven billion? Because of the thesis from FT zero-zero. Fine-tuning steers behavior; it does not inject knowledge. And that thesis is demonstrable on a one-billion model in minutes. You do not need a frontier model to learn that SFT changes format, that DPO shifts preference, or that abliteration deletes a direction. You need a model cheap enough to run the experiment ten times — to break it, reset it, and break it again. That is what teaching fine-tuning requires. This is the model the FT zero-zero lab loads, the one FT zero-eight's QLoRA walkthrough fine-tunes, and the one every VRAM calculation in FT zero-one uses as its worked example. It is not the smartest model in the family. It is the most teachable. [SLIDE 4 — The Ultra* datasets] The reason MiniCPM is the course's data reference, not just its model reference, is the Ultra-star family of open datasets. Three of them matter. UltraChat — large-scale multi-turn dialogue, the SFT dataset. UltraFeedback — preference and feedback pairs, the DPO dataset, introduced in the paper at arXiv twenty-three-ten point oh-one-three-seven-seven. And Ultra-FineWeb — a curated web pretraining mix, for continued pretraining or base distribution work. Each dataset maps to a steering layer. UltraChat to Layer three, SFT. UltraFeedback to Layer three, DPO. Ultra-FineWeb to Layer one, the base. And each appears in a specific course module — UltraChat in FT zero-four and FT twelve, UltraFeedback in FT zero-five and FT thirteen, Ultra-FineWeb in FT zero-six. The point is not that these are the best datasets. It is that they are open, documented, and reproducible. When FT thirteen teaches DPO, it teaches it on UltraFeedback because you can go read the dataset, find a concrete preference pair, and say: this is the signal DPO is optimizing. With a closed dataset, that pointer is impossible. Open data gives you a steering wheel you can inspect. That is why the data modules keep returning to Ultra-star. [SLIDE 5 — Why MiniCPM is the teaching vehicle] The Nature Communications paper — identifier s-four-one-four-six-seven dash oh-two-five dash six-one-oh-four-oh-five — makes the case formally. Three properties make MiniCPM the ideal teaching vehicle, and they map directly onto the course's needs. First, cheap iteration. A one-billion model fine-tunes in minutes on a consumer GPU. A student can run the full SFT, DPO, quantize, deploy loop in a single afternoon. You cannot teach fine-tuning on a model that takes six hours per run. Second, auditable provenance. Open weights, open data, open recipe. Every claim about what the model saw can be checked against the published corpus. This is the FT zero-two auditability predicate, made concrete on the base the course actually uses. Third, Apache-two-point-zero with no friction. No field-of-use restrictions, no monthly-active-user clauses, no named-product carve-outs — contrast Llama's Community License. A student can build on MiniCPM, ship a product, and never consult a lawyer about the model license. For a course that wants students to ship things, this matters. If you want a community companion to the formal paper, Sam Witteveen's walkthrough — at youtu dot be slash o-x-one-m-W-two-N-nine-Z-underscore-Y — covers the family overview and the fine-tuning workflow in a video format that complements this deep-dive. [SLIDE 6 — SWIFT vs LLaMA-Factory] Two frameworks dominate MiniCPM fine-tuning, and the course references both. The choice is a real engineering decision. SWIFT — Scalable lightWeight Infrastructure for Fine-Tuning — is ModelBest and OpenBMB's own framework. The first-party option. Its advantage is native, first-day support for every MiniCPM variant, including the vision and omni-modal models. Its disadvantage is that it is MiniCPM-flavored. It supports other models, but its center of gravity is the OpenBMB ecosystem. LLaMA-Factory is the general-purpose framework — a unified interface for fine-tuning dozens of model families. Its advantage is breadth. If your team already uses it for Llama, Qwen, and Mistral, adding MiniCPM is a config change. Its disadvantage is that support for the newest MiniCPM variants — especially the omni-modal models — can lag SWIFT by weeks to months. Here is the decision rule. If you are working with MiniCPM-V or MiniCPM-o, or you want first-party assurance that the chat template and modality handling are correct, use SWIFT. If you are fine-tuning MiniCPM five dash one B or three dash four B as part of a multi-model pipeline, use LLaMA-Factory. Both produce equivalent results on the text models. The difference is ecosystem fit. Neither is the framework this course uses for its core training-loop deep-dives — FTDD-zero-four covers TRL, FTDD-zero-five covers Axolotl, FTDD-zero-three covers Unsloth. SWIFT and LLaMA-Factory are the MiniCPM-specific entry points. [SLIDE 7 — Anti-patterns] Three anti-patterns to leave with. First, treating MiniCPM five dash one B as a production frontier model. It is not. It is a small, capable, open model — ideal for teaching, prototyping, and edge deployment. It will lose head-to-head capability comparisons against seventy-billion-plus models. The course uses it because it is cheap to iterate on, not because it is the strongest model available. Match the base to the deployment — that is FT zero-three. Second, assuming the Ultra-star datasets are clean because they are open. Open does not mean clean. UltraChat, UltraFeedback, and Ultra-FineWeb are open and documented, which means you can audit them — not that the audit has already been done. Run the same PII sweep, deduplication, and decontamination you would run on any dataset. Openness gives you the ability to vet. It does not do the vetting. Third, picking SWIFT or LLaMA-Factory without considering the modality. The text models fine-tune equivalently in both. The vision and omni-modal models do not. Choosing LLaMA-Factory for an omni-modal fine-tune because it is what you already use is a reasonable instinct that will cost you days of template-debugging. [SLIDE 8 — What you can now do] You can now place MiniCPM on the open spectrum and explain why it is the course's default base. You can map the family — five dash one B, three dash four B, V four point six, o four point five — to the modality gap each fills, and explain SigLip-four-hundred-M's role. You can trace the Ultra-star datasets to the steering layers they enable and cite the arXiv provenance. You can defend, with the Nature Communications paper, why an open-data family is the right teaching vehicle. And you can choose between SWIFT and LLaMA-Factory and defend the tradeoff. The lab asks you to load MiniCPM five dash one B, apply a format-steering LoRA via SWIFT, and measure the before-and-after behavior shift. The steering thesis, felt on the course's default base, in under thirty minutes. Next, deep-dive FTDD-zero-two: OLMo two and three plus Tulu three, from the Allen Institute. The fully-open comparison — the research-oriented counterpoint to MiniCPM's product-and-edge orientation. --- *End of deep-dive FTDD-01. Duration: approximately thirty minutes at one-hundred-forty words per minute.*