Add pipeline/rl/README.md
Browse files- pipeline/rl/README.md +11 -0
pipeline/rl/README.md
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# RL (GRPO/ PPO) sketch for tables
|
| 2 |
+
|
| 3 |
+
Rewards:
|
| 4 |
+
- JSON validity
|
| 5 |
+
- Exact field order
|
| 6 |
+
- Type checks per column
|
| 7 |
+
- Cell‑level F1 / edit distance
|
| 8 |
+
- Row count / primary‑key constraints
|
| 9 |
+
|
| 10 |
+
Loop: rollouts → score → GRPO step → checkpoint → eval.
|
| 11 |
+
Safety: reward clipping, degenerate‑mode detection, strict validators.
|