| --- |
| tags: |
| - classify-and-augment |
| - llm-annotated |
| --- |
| |
| # davanstrien/ledgar-top10-cascade-2tier |
|
|
| LLM-annotated dataset produced by [classify-and-augment](https://github.com/davanstrien/classify-and-augment). |
|
|
| ## Configuration |
|
|
| - **Model**: Cascade: `LiquidAI/LFM2.5-1.2B-Instruct` → `Qwen/Qwen3-4B-Instruct-2507` |
| - **Labels**: `Governing Laws`, `Notices`, `Counterparts`, `Entire Agreements`, `Severability`, `Survival`, `Amendments`, `Assignments`, `Expenses`, `Terms` |
| - **Input rows**: 200 |
| - **Output rows**: 200 |
|
|
| ## Label distribution |
|
|
| | Label | Real | Synthetic | Total | |
| |---|---:|---:|---:| |
| | `Amendments` | 17 | 0 | 17 | |
| | `Assignments` | 15 | 0 | 15 | |
| | `Counterparts` | 26 | 0 | 26 | |
| | `Entire Agreements` | 25 | 0 | 25 | |
| | `Expenses` | 14 | 0 | 14 | |
| | `Governing Laws` | 28 | 0 | 28 | |
| | `Notices` | 21 | 0 | 21 | |
| | `Severability` | 19 | 0 | 19 | |
| | `Survival` | 22 | 0 | 22 | |
| | `Terms` | 13 | 0 | 13 | |
|
|
| ## Cascade audit |
|
|
| | Tier | Model | Processed | Entropy | Parse fail % | `Governing Laws` count | `Notices` count | `Counterparts` count | `Entire Agreements` count | `Severability` count | `Survival` count | `Amendments` count | `Assignments` count | `Expenses` count | `Terms` count | |
| |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---| |
| | 1 | `LiquidAI/LFM2.5-1.2B-Instruct` | 200 | 0.000 | 0.0 | 0 | 0 | 200 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | |
| | 2 | `Qwen/Qwen3-4B-Instruct-2507` | 200 | 3.277 | 0.0 | 28 | 21 | 26 | 25 | 19 | 22 | 17 | 15 | 14 | 13 | |
|
|
| **Pairwise tier agreement:** |
|
|
| - Tier 1 vs Tier 2: 13.0% (26/200 examples) |
|
|
| **Escalation:** 174 examples (87.0%) had cross-tier disagreement; tier-3 arbitration not used (no tier-3 configured; tier-2 wins). |
|
|
| See [CASCADE_DESIGN.md](https://github.com/davanstrien/classify-and-augment/blob/main/CASCADE_DESIGN.md) for the disagreement-driven cascade strategy. |
|
|
|
|
