The minimum viable model probe · V1

Minimum Viable Model

Can scaffolding rescue a tiny model? We took eight sub-1B models and observed them across six prompting modes: raw (no scaffold), three orthogonal SCQ combinations, minimal typing, and minimal imports. The hypothesis: scaffolding should raise small-model scores meaningfully. The data did not confirm it.

top raw score: 3.40
top scaffolded: 3.00
scaffolding modes: 6

What each mode adds to the raw prompt

raw

Unmodified prompt

The baseline. The model sees only the user's question.

orthogonal_scq

Schema + constraints + question

Pre-fills the response schema, the boundary constraints, and reformats the question.

orthogonal_cq

Constraints + question

Constraints and reformatted question only.

orthogonal_sc

Schema + constraints

Schema and constraints only.

minimal_typing

Type hints inline

Provides minimum-viable type annotations on the expected output.

minimal_imports

Imports prepended

Pre-imports the libraries the model is expected to use.

Per-mode score table (0-10)

Rows sorted by raw mode score. A dash indicates the model did not complete the scaffolded variants in this run.

raw

granite4:350m: 3.4
qwen2.5-coder:0.5b: 3
qwen2.5:0.5b: 2.2
gemma3:1b: 2
smollm2:360m: 2
qwen3:0.6b: 0.4
qwen3.5:0.8b: 0
smollm2:135m: -0.2

8 entries · ordered by raw as observed in this run · not a ranking

Scaffolding mostly hurt

The headline finding is that in this dataset, scaffolded scores generally underperform raw for the same model. The strongest cell is qwen2.5-coder:0.5b + minimal_imports at 3.00 — matching raw, not exceeding it. The MVM hypothesis (that scaffolding lifts a tiny model to a usable level) is not supported by this sub-1B sample. This observation feeds the the code-generation service pipeline API's routing decisions: when a target quality threshold cannot be met by any (model, scaffolding) pair, the API escalates to a larger model.