"front"	"back"	"tags"
"Who is OpenBMB, and what does the name mean?"	"OpenBMB = 'Open Lab for Big Model Base' — a collaboration between Tsinghua University and the company ModelBest. The lab behind the MiniCPM family. All Apache-2.0, open weights + open data."	c3::ftdd01::recall
"Name the four MiniCPM family members and the modality each covers."	"(1) MiniCPM5-1B — ~1.08B dense text, the on-ramp hero. (2) MiniCPM3-4B — denser text, mid-size. (3) MiniCPM-V 4.6 — vision-language (pairs with SigLip-400M). (4) MiniCPM-o 4.5 — omni-modal, full-duplex (text+image+audio)."	c3::ftdd01::recall
"What are the headline specs of MiniCPM5-1B, and why does it matter to this course?"	"~1.08B parameters, released May 2026, Intelligence Index 17.9, ~2GB VRAM footprint. It is the course's DEFAULT on-ramp base — loaded in FT00's lab, fine-tuned in FT08's QLoRA walkthrough, and used as the worked example for every VRAM calc in FT01."	c3::ftdd01::recall
"Why 1B and not 7B as the teaching base?"	"The FT00 thesis (fine-tuning steers behavior, not knowledge) is DEMONSTRABLE on a 1B model in minutes. You don't need a frontier model to learn that SFT changes format or DPO shifts preference — you need a model cheap enough to run the experiment 10 times. 1B lets a student run the full SFT→DPO→quantize→deploy loop in an afternoon."	c3::ftdd01::analysis
"What is SigLip-400M, and why does it matter for MiniCPM-V?"	"SigLip-400M is a 400M-parameter image encoder. In MiniCPM-V 4.6, images pass through SigLip to become token-like embeddings, concatenated with text tokens, and processed by the shared transformer. It matters because it is itself open and well-understood — the vision-language stack is auditable down to the image encoder, not just the language backbone."	c3::ftdd01::recall
"Name the three Ultra* datasets and the steering layer each maps to."	"(1) UltraChat — multi-turn dialogue → Layer 3 SFT (FT04, FT12). (2) UltraFeedback — preference/feedback pairs → Layer 3 DPO (FT05, FT13). (3) Ultra-FineWeb — curated web pretraining mix → Layer 1 base / continued pretraining (FT06)."	c3::ftdd01::recall
"What is the arXiv citation for UltraFeedback, and why is it load-bearing?"	"arXiv:2310.01377 (Cui et al., 2023). It introduced UltraFeedback and established the open-preference-dataset pattern. Load-bearing because it means the course can point at a concrete preference pair and say 'this is the signal DPO is optimizing' — impossible with a closed dataset."	c3::ftdd01::recall
"Why are the Ultra* datasets the course's data reference, not just MiniCPM's model reference?"	"They are open, documented, and reproducible. When FT13 teaches DPO on UltraFeedback, you can go read the dataset and inspect the signal. A steering wheel you can inspect is a teaching tool a closed dataset cannot match. Open data = the ability to point at the exact signal being optimized."	c3::ftdd01::analysis
"State the three properties that make MiniCPM the ideal teaching vehicle."	"(1) CHEAP ITERATION — 1B fine-tunes in minutes on consumer GPU; full loop in an afternoon. (2) AUDITABLE PROVENANCE — open weights + data + recipe; point at every byte. (3) APACHE-2.0 — no MAU/field-of-use clauses; ship without license friction. Each maps to a concrete course need."	c3::ftdd01::application
"What is the Nature Communications paper identifier for MiniCPM, and what does it argue?"	"s41467-025-61040-5. It argues that a small, open, well-documented model is a viable research and teaching platform for the ENTIRE fine-tuning stack — walking through pretraining data curation, mid-training, and post-training with each stage's data and recipe disclosed."	c3::ftdd01::recall
"What is the Sam Witteveen MiniCPM walkthrough URL, and when would you use it?"	"youtu.be/ox1mW2N9Z_Y. The community companion video to the formal paper — covers the family overview and fine-tuning workflow in a video format that complements the Nature Comms paper. Use it when you want a walkthrough rather than a formal read."	c3::ftdd01::recall
"Compare SWIFT and LLaMA-Factory as MiniCPM fine-tuning frameworks."	"SWIFT = ModelBest/OpenBMB's FIRST-PARTY framework; native support for ALL MiniCPM variants (incl. vision/omni); first-party chat templates; OpenBMB-centric. LLaMA-Factory = GENERAL-PURPOSE; unified interface for dozens of families; add MiniCPM as a config change; vision/omni support lags SWIFT by weeks-months."	c3::ftdd01::application
"What is the decision rule for choosing SWIFT vs LLaMA-Factory?"	"MiniCPM-V or MiniCPM-o (or you want first-party modality assurance) → SWIFT. MiniCPM5-1B or 3-4B as part of a multi-model pipeline → LLaMA-Factory. Text models produce equivalent results in both; the difference is ecosystem fit and modality support."	c3::ftdd01::application
"Why is 'treating MiniCPM5-1B as a production frontier model' an anti-pattern?"	"It is a small, capable, OPEN model — ideal for teaching, prototyping, and edge deployment. It will lose head-to-head capability comparisons against 70B+ 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 (FT03)."	c3::ftdd01::analysis
"Why is 'Ultra* datasets are clean because they're open' an anti-pattern?"	"Open ≠ clean (the FT02 anti-pattern). UltraChat/UltraFeedback/Ultra-FineWeb are open and documented = you CAN audit them, not that the audit is done. Open datasets still need PII sweep, deduplication, and decontamination (FT06). Openness gives the ABILITY to vet; it does not do the vetting."	c3::ftdd01::analysis
"Why is MiniCPM's Apache-2.0 license specifically advantageous for a course that wants students to ship?"	"No field-of-use restrictions, no MAU 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, license friction is a teaching tax Apache-2.0 eliminates."	c3::ftdd01::analysis
"Fill in the modality progression: MiniCPM5-1B → ? → MiniCPM-V 4.6 → MiniCPM-o 4.5."	"MiniCPM5-1B (text) → MiniCPM3-4B (denser text) → MiniCPM-V 4.6 (+ SigLip-400M, vision-language) → MiniCPM-o 4.5 (+ audio, full-duplex omni). Each step adds a modality or capability the previous lacked; SigLip-400M is the component that turns a text base into a vision-language model."	c3::ftdd01::application