model card: final results

README.md

@@ -13,19 +13,17 @@ tags:
 
 # tiny-aya-global-sql-selfheal-bco (BCO, pairs only)
 
-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 detected by executing SQL on Spider databases, classified, corrected by
-Aya Expanse 32B (fixes verified by execution; healer never sees gold SQL), and
-retrained through AgentTune.
+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:** BCO (AgentTune `RetrainConfig`, TRL), same LoRA config, 153 verified fixes (label=True) + 153 original wrong queries (label=False).
+Recipe: BCO (AgentTune RetrainConfig, TRL), same LoRA config, 153 verified fixes (label True) + 153 original wrong queries (label False).
 
-**Result:** Regressed on the gate split; superseded by the +replay variant.
-AgentTune's DeploymentGate **blocked deployment** (task accuracy regressed) — the
-gate doing exactly what PS #8 requires. Full analysis (likelihood displacement on
-minimal-edit SQL pairs + catastrophic forgetting) in the article.
+Result: Regressed on the gate split and was BLOCKED; superseded by the +replay variant.
 
 - Project page + code: https://huggingface.co/spaces/HideIron/selfheal-sql
-- All run logs / gate decisions: https://huggingface.co/datasets/HideIron/selfheal-sql-runs
-- Training data: 158 execution-verified fix pairs from `heal/baseline-pool600` in the runs repo
+- 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)