Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags: [clinical, emergency-medicine, self-play, rl, ces-zero]
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
# maple-signal-64 — CES-Zero cleanv3 step-162 ("s162")
|
| 7 |
+
|
| 8 |
+
Self-play RL checkpoint for clinical ED simulation. **First checkpoint in the CES-Zero
|
| 9 |
+
project to finish at-or-above its SFT seed on the paired 120-case official eval**
|
| 10 |
+
(0.8646 single-draw vs seed 0.8583; MEAN of two draws ≈0.8724; two-draw oracle 0.9038).
|
| 11 |
+
|
| 12 |
+
- Seed: ck20 (canonical 0.8769 SFT checkpoint), trained with seq-level PPO clip
|
| 13 |
+
(CLIP_EPS 0.01/0.015, ratio vs seed) + KL 0.04, lr 1e-5, one offline self-play cycle
|
| 14 |
+
(648 samples, husk-purged corpus).
|
| 15 |
+
- Rung ladder: step064 0.8487 -> **step162 0.8646** -> final 0.8369 (select checkpoints
|
| 16 |
+
by full-120 eval; the peak is mid-trajectory).
|
| 17 |
+
- Full reproducibility freeze (trainer, corpus, eval pipeline, scoring lane, results):
|
| 18 |
+
branch **`s162`** of github.com/monk1337/cesmainrepo (`s162_freeze/README_S162.md`).
|
| 19 |
+
- Siblings: pebble-orbit-31 (step064), quartz-harbor-19 (final), willow-comet-58
|
| 20 |
+
(shuffled-advantage control), aurora-kite-77 (patient model).
|
| 21 |
+
|
| 22 |
+
Serve (vllm >= 0.27): `vllm serve models4world/maple-signal-64 --dtype bfloat16
|
| 23 |
+
--enable-auto-tool-choice --tool-call-parser qwen3_xml --reasoning-parser qwen3
|
| 24 |
+
--language-model-only`
|