| --- |
| base_model: CohereLabs/tiny-aya-global |
| library_name: peft |
| license: cc-by-nc-4.0 |
| datasets: |
| - HideIron/selfheal-sql-spider |
| tags: |
| - text-to-sql |
| - lora |
| - self-healing |
| - agenttune |
| --- |
| |
| # tiny-aya-global-sql-selfheal (DPO) |
|
|
| LoRA adapter for CohereLabs/tiny-aya-global (3.35B) from a fully automated |
| self-healing text-to-SQL loop (Lexsi Labs Hackathon, Track 3, PS #8). Failures are |
| detected by executing the SQL on real Spider databases, classified, corrected by |
| Aya Expanse 32B (every fix verified by execution; the healer never sees the gold |
| query), and retrained through AgentTune. Deployment is decided by AgentTune's |
| DeploymentGate on held-out data. |
|
|
| Recipe: DPO (AgentTune RetrainConfig, TRL), LoRA r=16/alpha=32 on all projections, 158 verified pairs, 3 epochs, lr 1e-5, beta 0.1. |
|
|
| Result: Gate split 58.0% vs baseline 62.0% (-4pp, fixed 4 / broke 8); test 44.6% vs 45.6%. The DeploymentGate BLOCKED this adapter (task accuracy regressed). The write-up explains why: likelihood displacement on minimal-edit SQL pairs plus catastrophic forgetting. |
|
|
| - Project page + code: https://huggingface.co/spaces/HideIron/selfheal-sql |
| - All run logs and gate decisions: https://huggingface.co/datasets/HideIron/selfheal-sql-runs |
| - Training data: 158 execution-verified fix pairs (heal/baseline-pool600 in the runs repo) |
|
|
