Visual extraction and scoring observation: local MiniCPM-V 4.5 and 4.6 aliases · V3

Model observation · V3

Sigma5C observed local MiniCPM-V aliases on image and text fixtures in CPU-only environments. Task-specific checks sometimes assigned different scores from a substring-presence check on the same responses, making the scoring method part of the observation.

Why this observation matters

Visual extraction workflows depend on what is checked as well as what a model writes. A response containing an expected value can still leave its meaning or association unresolved, so scoring boundaries matter for interpreting the record.

CEO

Scope before application

The observation makes the limits of a visual extraction record visible.

CFO

Interpret the metric

A scoring-method difference is an assurance question, not a financial or business outcome.

CTO

Inspect the checks

The retained records separate execution state, scorer behavior and untested associations.

What we did

We examined the June 29 records from observation/scoring version V3, which is a protocol label rather than a model version. Two local model aliases were observed across three CPU-only environments. One cohort contains 48 responses to the original fixture battery, scored both by substring presence and task-specific checks. A separate native-lane cohort contains 18 responses. The two cohorts have different fixtures and are kept separate; no pooled model score is presented.

Selected observations under these conditions

The scoring method changes the reading

The two scoring methods agreed on some responses and assigned different numeric values to others.

Data teams can inspect which response properties a metric actually measures before applying it to extraction workflows.

The chart compares scorers on the same June 29 responses; it does not measure a change between different model runs.

Completion and content checks differ

Both the original-battery cohort and the separate native-lane cohort recorded completed calls.

Execution status and content assessment can be retained as distinct fields in an internal review.

Completed calls do not establish complete or correctly associated field extraction.

Task-specific checks retain limits

The retained scoring protocol leaves some diagram direction and multi-image value association outside its checks.

A task-specific scorer still needs a declared coverage boundary.

The records support descriptive findings for these fixtures; they do not establish general visual or video understanding.

Selected recorded aggregates

How the two scoring methods differed on the same responses

Same numeric score: 37 records
Task-specific score lower: 5 records
Task-specific score higher: 6 records

Source: June 29, 2026 UTC MiniCPM-V V3 original-battery response and scoring records. 48 responses from two local aliases across three CPU-only environments. Each response received the legacy substring check and the task-specific scorer. The axis is 0–48 records. The chart measures scoring-method disagreement, not the number of correct answers; categories partition the same 48 responses.

Declared conditions

June 29, 2026 UTC. Local Ollama requests used the recorded aliases minicpm-v4.5:latest and minicpm-v4.6:latest, CPU-only execution, bounded context and output, fixed generated fixtures, and deterministic task-specific scoring. Processor snapshots matched the requested models with no recorded VRAM allocation. These local aliases identify the observed artifacts; the label 4.6 must not be interpreted as an observation of every model size in that release family.

Potential uses

CEO / strategy

The records could provide dated, task-bound context for a visual extraction review.

CFO / oversight

The evidence could support an assurance review that separates completion from the scope of content checks.

CTO / data teams

Scoring differences, negative controls and per-response evidence could inform extraction validation design.

What's in the data

The retained evidence contains the following record classes. This description does not assert that every artifact is currently released through the API.

Separate original-battery and native-lane records with model aliases, timestamps and CPU-only processor evidence.
Task-specific scores, legacy substring values and their per-response differences.
Fixture-integrity receipts, request settings, timeout policies and token/timing telemetry.
Declared scorer residuals and earlier observation/scoring history, retained separately.