root card

README.md

@@ -0,0 +1,95 @@
+---
+license: apache-2.0
+tags: [chess, reinforcement-learning, grpo]
+---
+
+# Chess-RL-Models
+
+RL (GRPO) post-training checkpoint trajectories for the chess pre-to-post
+compute-allocation study. Companion to
+[Chess-Pretrain-Models](https://huggingface.co/pavelslab-nyu/Chess-Pretrain-Models)
+(pretraining bases) and
+[Chess-SFT-Models](https://huggingface.co/pavelslab-nyu/Chess-SFT-Models)
+(SFT initialisations). Model names match across all three repos, so
+`model_50m_17B` here is the RL run started from `model_50m_17B` there.
+
+**28 models, 680 checkpoints.** Steps kept: every step ≡ 0 mod 100,
+plus each run's final step.
+
+## Layout
+
+```
+model_{size}_{pretraining_tokens}/
+└── global_step_{N}/
+    ├── config.json
+    ├── model.safetensors
+    ├── tokenizer.py      # custom tokenizer -> trust_remote_code=True
+    ├── vocab.json
+    └── ...
+```
+
+## Loading
+
+The remote-code resolver ignores `subfolder=`, so download the step folder and
+load the **local path** (each step folder is self-contained):
+
+```python
+from huggingface_hub import snapshot_download
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+name, step = "model_50m_17B", "global_step_2000"
+p = snapshot_download("Pre2Post-Chess-RL/Chess-RL-Models", allow_patterns=f"{name}/{step}/*") + f"/{name}/{step}"
+model = AutoModelForCausalLM.from_pretrained(p, trust_remote_code=True)
+tok = AutoTokenizer.from_pretrained(p, trust_remote_code=True)
+```
+
+## Models
+
+| model | size | pretrain tokens | compute | alpha | beta | ckpts | steps | run id |
+|---|---|---|---|---|---|---|---|---|
+| `model_200m_0.27B` | 200m | 0.267B | 6p5e18 | 0.05 | 0.1 | 10 | 100–1000 | `C6p5e18_200m_alpha0.050_beta0.100` |
+| `model_200m_1.1B` | 200m | 1.07B | 6p5e18 | 0.2 | 0.1 | 20 | 100–2000 | `C6p5e18_200m_alpha0.200_beta0.100` |
+| `model_200m_2.1B` | 200m | 2.13B | 6p5e18 | 0.4 | 0.1 | 20 | 100–2000 | `C6p5e18_200m_alpha0.400_beta0.100` |
+| `model_200m_4.0B` | 200m | 4B | 6p5e18 | 0.75 | 0.1 | 10 | 100–1000 | `C6p5e18_200m_alpha0.750_beta0.100` |
+| `model_200m_5.3B` | 200m | 5.33B | 6p5e18 | 1.0 | 0.1 | 20 | 100–2000 | `C6p5e18_200m_alpha1.000_beta0.100` |
+| `model_20m_0.53B` | 20m | 0.527B | 6p5e18 | 0.01 | 0.008 | 20 | 100–2000 | `C6p5e18_20m_alpha0.010_beta0.008` |
+| `model_20m_1.6B` | 20m | 1.58B | 6p5e18 | 0.03 | 0.008 | 50 | 100–5000 | `C6p5e18_20m_alpha0.030_beta0.008` |
+| `model_20m_2.6B` | 20m | 2.64B | 6p5e18 | 0.05 | 0.008 | 49 | 100–4900 | `C6p5e18_20m_alpha0.050_beta0.008` |
+| `model_20m_5.3B` | 20m | 5.27B | 6p5e18 | 0.1 | 0.008 | 50 | 100–5000 | `C6p5e18_20m_alpha0.100_beta0.008` |
+| `model_20m_11B` | 20m | 10.5B | 6p5e18 | 0.2 | 0.008 | 50 | 100–5000 | `C6p5e18_20m_alpha0.200_beta0.008` |
+| `model_20m_16B` | 20m | 15.8B | 6p5e18 | 0.3 | 0.008 | 30 | 100–3000 | `C6p5e18_20m_alpha0.300_beta0.008` |
+| `model_20m_21B` | 20m | 21.1B | 6p5e18 | 0.4 | 0.008 | 50 | 100–5000 | `C6p5e18_20m_alpha0.400_beta0.008` |
+| `model_20m_32B` | 20m | 31.6B | 6p5e18 | 0.6 | 0.008 | 50 | 100–5000 | `C6p5e18_20m_alpha0.600_beta0.008` |
+| `model_20m_40B` | 20m | 39.6B | 6p5e18 | 0.75 | 0.008 | 49 | 100–4900 | `C6p5e18_20m_alpha0.750_beta0.008` |
+| `model_20m_53B` | 20m | 52.7B | 6p5e18 | 1.0 | 0.008 | 32 | 100–5000 | `C6p5e18_20m_alpha1.000_beta0.008` |
+| `model_50m_0.23B` | 50m | 0.229B | 6p5e18 | 0.01 | 0.023 | 20 | 100–2000 | `C6p5e18_50m_alpha0.010_beta0.023` |
+| `model_50m_0.69B` | 50m | 0.687B | 6p5e18 | 0.03 | 0.023 | 12 | 100–1200 | `C6p5e18_50m_alpha0.030_beta0.023` |
+| `model_50m_1.1B` | 50m | 1.15B | 6p5e18 | 0.05 | 0.023 | 10 | 100–1000 | `C6p5e18_50m_alpha0.050_beta0.023` |
+| `model_50m_2.3B` | 50m | 2.29B | 6p5e18 | 0.1 | 0.023 | 20 | 100–2000 | `C6p5e18_50m_alpha0.100_beta0.023` |
+| `model_50m_4.6B` | 50m | 4.58B | 6p5e18 | 0.2 | 0.023 | 20 | 100–2000 | `C6p5e18_50m_alpha0.200_beta0.023` |
+| `model_50m_9.2B` | 50m | 9.16B | 6p5e18 | 0.4 | 0.023 | 20 | 100–2000 | `C6p5e18_50m_alpha0.400_beta0.023` |
+| `model_50m_17B` | 50m | 17.2B | 6p5e18 | 0.75 | 0.023 | 20 | 100–2000 | `C6p5e18_50m_alpha0.750_beta0.023` |
+| `model_50m_23B` | 50m | 22.9B | 6p5e18 | 1.0 | 0.023 | 10 | 100–1000 | `C6p5e18_50m_alpha1.000_beta0.023` |
+| `model_50m_41B` | 50m | 41.2B | 6p5e19 | 0.18 | 0.002 | 18 | 100–2000 | `C6p5e19_50m_alpha0.180_beta0.002` |
+| `model_680m_0.32B` | 680m | 0.319B | 6p5e18 | 0.2 | 0.296 | 5 | 100–500 | `C6p5e18_680m_alpha0.200_beta0.296` |
+| `model_680m_1.6B` | 680m | 1.6B | 6p5e18 | 1.0 | 0.296 | 1 | 2000–2000 | `C6p5e18_680m_alpha1.000_beta0.296` |
+| `model_680m_6.4B` | 680m | 6.4B | 6p5e19 | 0.4 | 0.03 | 10 | 100–1000 | `C6p5e19_680m_alpha0.400_beta0.030` |
+| `model_680m_16B` | 680m | 16B | 6p5e19 | 1.0 | 0.03 | 4 | 1200–1900 | `C6p5e19_680m_alpha1.000_beta0.030` |
+
+## Not included
+
+These runs exist in the study but had no usable HF-format weights on disk at
+upload time (evals survive, the weights were cleaned up):
+
+- `C6p5e18_200m_alpha0.100_beta0.100` — no usable checkpoint on disk
+- `C6p5e18_32m_alpha0.200_beta0.013` — no usable checkpoint on disk
+- `C6p5e18_32m_alpha0.400_beta0.013` — not in run_metadata.csv - cannot derive name
+- `C6p5e18_680m_alpha0.400_beta0.296` — no usable checkpoint on disk
+- `C6p5e18_680m_alpha0.750_beta0.296` — no usable checkpoint on disk
+- `C6p5e19_200m_alpha0.200_beta0.007` — no usable checkpoint on disk
+- `C6p5e19_200m_alpha0.400_beta0.007` — no usable checkpoint on disk
+- `C6p5e19_200m_alpha0.750_beta0.007` — no usable checkpoint on disk
+- `C6p5e19_200m_alpha1.000_beta0.007` — no usable checkpoint on disk
+- `C6p5e19_680m_alpha0.200_beta0.030` — no usable checkpoint on disk
+- `C6p5e19_680m_alpha0.750_beta0.030` — no usable checkpoint on disk
+- `C6p5e19_680m_alpha2.000_beta0.030` — no usable checkpoint on disk