Module FT03 — Base Model Selection

Base Model Selection

The base selection rubric — task, hardware, license, openness, ecosystem — weighted by use case. Why MiniCPM5-1B is the course default, when to graduate, and which properties matter (and which people over-index on).

minutes

artifacts

sub-sections

Base selection is not a leaderboard lookup. It is a constrained optimization across five dimensions whose weights shift by use case: a phone assistant weights hardware near-absolutely; a HIPAA bot weights openness as a hard gate; a pentest tool weights the ablation ecosystem. The most expensive mistakes — license failure, provenance loss, capability ceiling — are made here, before a single token is trained.

Key Claims

Load-Bearing Claims

Base selection is a five-dimension rubric — task, hardware, license, openness, ecosystem — and the weights shift by use case, not hold fixed. Task sets the capability floor; hardware sets the size ceiling (FT01); license and openness are gates that can veto; ecosystem is the friction coefficient.

MiniCPM5-1B is the course default because it wins on the dimensions that matter for learning: ~1B params for fast iteration, Apache-2.0 license, open-data pipeline (auditable, FT02), and first-class ecosystem support. You graduate to a larger base only when a signal fires — never 'just because bigger.'

The properties that matter for fine-tuning are tokenizer/domain fit, context length, chat template quality, license, open-data availability, and ecosystem support. The properties people over-index on — raw benchmark scores and parameter count in isolation — are the wrong frame: you are steering a base, not competing on a leaderboard, and the FT01 VRAM math decides what fits.

Default to the instruct checkpoint. Use base when you want to build behavior from scratch. Use chat when you want to preserve existing alignment and steer narrowly (abliteration). The wrong-base anti-patterns — instruct-when-you-wanted-base, merged-community-no-provenance, base-too-small, ignoring-tokenizer-fit — are made upstream of training and are the most expensive to fix.

After This Module

State the five-dimension base selection rubric (task, hardware, license, openness, ecosystem) and explain why the weights shift by use case rather than holding fixed.

Defend the choice of MiniCPM5-1B as the course's default teaching base, and name the signals that tell you to graduate to a larger base.

Identify the base-model properties that genuinely matter for fine-tuning versus those that people over-index on (raw benchmarks, parameter count in isolation).

Place the major base families (OpenBMB MiniCPM, Qwen, Llama, DeepSeek, Mistral/Mixtral, SmolLM3/OLMo/Tülu) on a size × openness map and describe each family's fine-tuning profile.

Decide, for a given goal, whether to start from a base, instruct, or chat checkpoint — and recognize the wrong-base anti-patterns.

Artifacts

Teaching Document

~3,200 words; 6 sub-sections — the five-dimension rubric, why MiniCPM5-1B is the course default, properties that matter (and don't), the major base families, base vs instruct vs chat, wrong-base anti-patterns

READ

Diagrams

5 Mermaid diagrams — the five-dimension rubric, the base selection decision tree, base families on a size × openness grid, base/instruct/chat decision flowchart, the MiniCPM on-ramp → graduation progression

READ

Slide Deck

reveal.js slide deck, dark theme, design-system teal

READ

Teaching Script