Initialize reviewed checkpoint archive

README.md

@@ -0,0 +1,35 @@
+---
+library_name: transformers
+tags:
+- policy-gradient
+- adaptive-group-size
+---
+
+# Adaptive group-size policy checkpoints
+
+Final fixed-group and adaptive policy checkpoints for GSM8K, mathematical
+reasoning and code generation. This is a multi-model archive: choose a model
+subdirectory from `checkpoints.json`, rather than loading the repository root.
+
+| Family | Models | Final step |
+|---|---:|---:|
+| GSM8K / Qwen2.5-1.5B-Instruct | 21 | 500 |
+| Math / Qwen3-8B-SFT496 | 6 | 300 |
+| Math / Qwen2.5-Math-7B | 6 | 500 |
+| Coding / DeepSeek-Coder-1.3B-Base | 6 | 500 |
+
+The manifest lists planned models; a model is available after its folder has
+been committed. Historical hybrid runs retain their original configuration
+name. Base-model licenses remain applicable; no optimizer states or training
+datasets are published here.
+
+```python
+from pathlib import Path
+from huggingface_hub import snapshot_download
+from transformers import AutoModelForCausalLM, AutoTokenizer
+subdir = "models/gsm8k/fixed2/seed42/step500"
+root = snapshot_download("Gsj49/adaptiveG", allow_patterns=[subdir + "/*"])
+folder = Path(root) / subdir
+tokenizer = AutoTokenizer.from_pretrained(folder)
+model = AutoModelForCausalLM.from_pretrained(folder)
+```

checkpoints.json

@@ -0,0 +1,434 @@
+{
+  "repository": "Gsj49/adaptiveG",
+  "models": [
+    {
+      "id": "coding-fixed16",
+      "family": "coding",
+      "model": "DeepSeek-Coder-1.3B-Base",
+      "method": "fixed16",
+      "seed": 42,
+      "step": 500,
+      "path": "models/coding/fixed16/seed42/step500",
+      "base_model": "deepseek-ai/deepseek-coder-1.3b-base",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "coding-fixed32",
+      "family": "coding",
+      "model": "DeepSeek-Coder-1.3B-Base",
+      "method": "fixed32",
+      "seed": 42,
+      "step": 500,
+      "path": "models/coding/fixed32/seed42/step500",
+      "base_model": "deepseek-ai/deepseek-coder-1.3b-base",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "coding-fixed4",
+      "family": "coding",
+      "model": "DeepSeek-Coder-1.3B-Base",
+      "method": "fixed4",
+      "seed": 42,
+      "step": 500,
+      "path": "models/coding/fixed4/seed42/step500",
+      "base_model": "deepseek-ai/deepseek-coder-1.3b-base",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "coding-fixed64",
+      "family": "coding",
+      "model": "DeepSeek-Coder-1.3B-Base",
+      "method": "fixed64",
+      "seed": 42,
+      "step": 500,
+      "path": "models/coding/fixed64/seed42/step500",
+      "base_model": "deepseek-ai/deepseek-coder-1.3b-base",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "coding-fixed8",
+      "family": "coding",
+      "model": "DeepSeek-Coder-1.3B-Base",
+      "method": "fixed8",
+      "seed": 42,
+      "step": 500,
+      "path": "models/coding/fixed8/seed42/step500",
+      "base_model": "deepseek-ai/deepseek-coder-1.3b-base",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "coding-hybrid",
+      "family": "coding",
+      "model": "DeepSeek-Coder-1.3B-Base",
+      "method": "hybrid",
+      "seed": 42,
+      "step": 500,
+      "path": "models/coding/hybrid/seed42/step500",
+      "base_model": "deepseek-ai/deepseek-coder-1.3b-base",
+      "method_note": "Historical hybrid configuration; retained under its original name."
+    },
+    {
+      "id": "qwen25-math-g16",
+      "family": "math_qwen25",
+      "model": "Qwen2.5-Math-7B",
+      "method": "g16",
+      "seed": 42,
+      "step": 500,
+      "path": "models/math_qwen25/g16/seed42/step500",
+      "base_model": "Qwen/Qwen2.5-Math-7B",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "qwen25-math-g32",
+      "family": "math_qwen25",
+      "model": "Qwen2.5-Math-7B",
+      "method": "g32",
+      "seed": 42,
+      "step": 500,
+      "path": "models/math_qwen25/g32/seed42/step500",
+      "base_model": "Qwen/Qwen2.5-Math-7B",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "qwen25-math-g4",
+      "family": "math_qwen25",
+      "model": "Qwen2.5-Math-7B",
+      "method": "g4",
+      "seed": 42,
+      "step": 500,
+      "path": "models/math_qwen25/g4/seed42/step500",
+      "base_model": "Qwen/Qwen2.5-Math-7B",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "qwen25-math-g64",
+      "family": "math_qwen25",
+      "model": "Qwen2.5-Math-7B",
+      "method": "g64",
+      "seed": 42,
+      "step": 500,
+      "path": "models/math_qwen25/g64/seed42/step500",
+      "base_model": "Qwen/Qwen2.5-Math-7B",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "qwen25-math-g8",
+      "family": "math_qwen25",
+      "model": "Qwen2.5-Math-7B",
+      "method": "g8",
+      "seed": 42,
+      "step": 500,
+      "path": "models/math_qwen25/g8/seed42/step500",
+      "base_model": "Qwen/Qwen2.5-Math-7B",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "qwen25-math-hybrid",
+      "family": "math_qwen25",
+      "model": "Qwen2.5-Math-7B",
+      "method": "hybrid",
+      "seed": 42,
+      "step": 500,
+      "path": "models/math_qwen25/hybrid/seed42/step500",
+      "base_model": "Qwen/Qwen2.5-Math-7B",
+      "method_note": "Historical hybrid configuration; retained under its original name."
+    },
+    {
+      "id": "qwen3-math-fixed16",
+      "family": "math_qwen3",
+      "model": "Qwen3-8B-SFT496",
+      "method": "fixed16",
+      "seed": 42,
+      "step": 300,
+      "path": "models/math_qwen3/fixed16/seed42/step300",
+      "base_model": "Qwen/Qwen3-8B",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "qwen3-math-fixed32",
+      "family": "math_qwen3",
+      "model": "Qwen3-8B-SFT496",
+      "method": "fixed32",
+      "seed": 42,
+      "step": 300,
+      "path": "models/math_qwen3/fixed32/seed42/step300",
+      "base_model": "Qwen/Qwen3-8B",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "qwen3-math-fixed4",
+      "family": "math_qwen3",
+      "model": "Qwen3-8B-SFT496",
+      "method": "fixed4",
+      "seed": 42,
+      "step": 300,
+      "path": "models/math_qwen3/fixed4/seed42/step300",
+      "base_model": "Qwen/Qwen3-8B",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "qwen3-math-fixed64",
+      "family": "math_qwen3",
+      "model": "Qwen3-8B-SFT496",
+      "method": "fixed64",
+      "seed": 42,
+      "step": 300,
+      "path": "models/math_qwen3/fixed64/seed42/step300",
+      "base_model": "Qwen/Qwen3-8B",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "qwen3-math-fixed8",
+      "family": "math_qwen3",
+      "model": "Qwen3-8B-SFT496",
+      "method": "fixed8",
+      "seed": 42,
+      "step": 300,
+      "path": "models/math_qwen3/fixed8/seed42/step300",
+      "base_model": "Qwen/Qwen3-8B",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "qwen3-math-hybrid",
+      "family": "math_qwen3",
+      "model": "Qwen3-8B-SFT496",
+      "method": "hybrid",
+      "seed": 42,
+      "step": 300,
+      "path": "models/math_qwen3/hybrid/seed42/step300",
+      "base_model": "Qwen/Qwen3-8B",
+      "method_note": "Historical hybrid configuration; retained under its original name."
+    },
+    {
+      "id": "gsm8k-fixed16-s42",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed16",
+      "seed": 42,
+      "step": 500,
+      "path": "models/gsm8k/fixed16/seed42/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed16-s43",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed16",
+      "seed": 43,
+      "step": 500,
+      "path": "models/gsm8k/fixed16/seed43/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed16-s44",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed16",
+      "seed": 44,
+      "step": 500,
+      "path": "models/gsm8k/fixed16/seed44/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed2-s42",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed2",
+      "seed": 42,
+      "step": 500,
+      "path": "models/gsm8k/fixed2/seed42/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed2-s43",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed2",
+      "seed": 43,
+      "step": 500,
+      "path": "models/gsm8k/fixed2/seed43/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed2-s44",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed2",
+      "seed": 44,
+      "step": 500,
+      "path": "models/gsm8k/fixed2/seed44/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed32-s42",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed32",
+      "seed": 42,
+      "step": 500,
+      "path": "models/gsm8k/fixed32/seed42/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed32-s43",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed32",
+      "seed": 43,
+      "step": 500,
+      "path": "models/gsm8k/fixed32/seed43/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed32-s44",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed32",
+      "seed": 44,
+      "step": 500,
+      "path": "models/gsm8k/fixed32/seed44/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed4-s42",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed4",
+      "seed": 42,
+      "step": 500,
+      "path": "models/gsm8k/fixed4/seed42/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed4-s43",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed4",
+      "seed": 43,
+      "step": 500,
+      "path": "models/gsm8k/fixed4/seed43/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed4-s44",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed4",
+      "seed": 44,
+      "step": 500,
+      "path": "models/gsm8k/fixed4/seed44/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed64-s42",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed64",
+      "seed": 42,
+      "step": 500,
+      "path": "models/gsm8k/fixed64/seed42/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed64-s43",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed64",
+      "seed": 43,
+      "step": 500,
+      "path": "models/gsm8k/fixed64/seed43/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed64-s44",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed64",
+      "seed": 44,
+      "step": 500,
+      "path": "models/gsm8k/fixed64/seed44/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed8-s42",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed8",
+      "seed": 42,
+      "step": 500,
+      "path": "models/gsm8k/fixed8/seed42/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed8-s43",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed8",
+      "seed": 43,
+      "step": 500,
+      "path": "models/gsm8k/fixed8/seed43/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-fixed8-s44",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "fixed8",
+      "seed": 44,
+      "step": 500,
+      "path": "models/gsm8k/fixed8/seed44/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Fixed group-size training."
+    },
+    {
+      "id": "gsm8k-hybrid-s42",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "hybrid",
+      "seed": 42,
+      "step": 500,
+      "path": "models/gsm8k/hybrid/seed42/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Historical hybrid configuration; retained under its original name."
+    },
+    {
+      "id": "gsm8k-hybrid-s43",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "hybrid",
+      "seed": 43,
+      "step": 500,
+      "path": "models/gsm8k/hybrid/seed43/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Historical hybrid configuration; retained under its original name."
+    },
+    {
+      "id": "gsm8k-hybrid-s44",
+      "family": "gsm8k",
+      "model": "Qwen2.5-1.5B-Instruct",
+      "method": "hybrid",
+      "seed": 44,
+      "step": 500,
+      "path": "models/gsm8k/hybrid/seed44/step500",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "method_note": "Historical hybrid configuration; retained under its original name."
+    }
+  ]
+}