Instruction file imported from sam4rano/yoruba_ocr_research (
.cursor/rules/03-results-reporting.mdc). Copyright stays with the author.
Skill: Results Reporting
Lead with findings, not setups
GOOD: "Fine-tuning reduces DER by 47% relative to the strongest baseline..." BAD: "Table 2 shows the results of our experiment comparing..."
Every number rule
State value -> (Table/Figure reference) -> interpretation
When referring to the supervised result, name PaddleOCR-VL-1.6 SFT and
state its recorded train scope (lm_head, non_vision, or all). Never insert
example metric values: every value must come from the regenerated metrics CSV.
Standard results table format
| Model | CER ↓ | WER ↓ | DER ↓ | Notes |
|---|---|---|---|---|
| Bold best result per column. | ||||
| Underline second-best. | ||||
| Report to 1 decimal place. |
Ablation paragraph structure
- What was removed/changed
- Quantitative delta
- Interpretation: what this reveals about the model
- One sentence connecting back to the main finding
Zero-shot vs supervised analysis
Compare CER, WER, and DER separately. Treat any claim about semantic recovery or tonal fidelity as a result to be established by the regenerated data, not a prewritten conclusion.
Language for uncertainty
"suggests" not "proves" "we observe" not "we demonstrate" "consistent with" not "confirms"