Upload folder using huggingface_hub

src/mathcompose/train/colab_train.ipynb

@@ -7,9 +7,7 @@
     "# mathcompose \u2014 QLoRA training on Colab (free T4)\n",
     "\n",
     "Pulls the code **and** the Model-V dataset from HF Hub with just your token \u2014 no\n",
-    "GitHub needed (run `python scripts/publish_hf.py` first to publish them).\n",
-    "\n",
-    "**Runtime \u2192 Change runtime type \u2192 T4 GPU** before running."
+    "GitHub needed. **Runtime \u2192 Change runtime type \u2192 T4 GPU** before running."
    ]
   },
   {
@@ -18,7 +16,7 @@
    "execution_count": null,
    "outputs": [],
    "source": [
-    "# 1. GPU deps (pinned to the versions this project was built against) + 4-bit kernels\n",
+    "# 1. GPU deps (versions this project was built against) + 4-bit kernels\n",
     "!pip -q install \"transformers>=5.13,<6\" \"trl>=1.7,<2\" \"peft>=0.19,<0.20\" \\\n",
     "    \"accelerate>=1.14\" \"datasets>=5.0\" \"huggingface_hub>=1.22\" sympy bitsandbytes"
    ]
@@ -32,11 +30,10 @@
     "# 2. Token -> pull the code+data snapshot -> install\n",
     "import os\n",
     "from huggingface_hub import HfApi, snapshot_download, login\n",
-    "os.environ[\"HF_TOKEN\"] = \"hf_xxx\"   # paste, or use Colab Secrets: userdata.get('HF_TOKEN')\n",
+    "os.environ[\"HF_TOKEN\"] = \"hf_xxx\"   # paste, or Colab Secrets: userdata.get('HF_TOKEN')\n",
     "login(os.environ[\"HF_TOKEN\"])\n",
     "USER = HfApi().whoami()[\"name\"]\n",
-    "CODE_REPO = f\"{USER}/mathcompose\"\n",
-    "proj = snapshot_download(CODE_REPO, repo_type=\"model\")\n",
+    "proj = snapshot_download(f\"{USER}/mathcompose\", repo_type=\"model\")\n",
     "%cd {proj}\n",
     "!pip -q install -e .\n",
     "print(\"project at\", proj, \"| data:\", os.listdir(\"data/verifier\"))"
@@ -48,7 +45,7 @@
    "execution_count": null,
    "outputs": [],
    "source": [
-    "# 3. Sanity: GPU present?\n",
+    "# 3. GPU present?\n",
     "import torch; print(\"CUDA:\", torch.cuda.is_available(), torch.cuda.get_device_name(0) if torch.cuda.is_available() else \"\")"
    ]
   },
@@ -72,8 +69,8 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "## 5. Eval V on ProcessBench \u2014 base vs tuned (Maj@8 headline) + PRMBench\n",
-    "Start with `--limit` to get a fast read, then drop it for the full 3,400-example score."
+    "## 5. Eval V on ProcessBench \u2014 base vs tuned (Maj@8) + PRMBench\n",
+    "Use `--limit` for a fast read; drop it for the full 3,400-example score."
    ]
   },
   {
@@ -85,14 +82,14 @@
     "!python -m mathcompose.eval.run processbench --tag base  --maj-k 1 --limit 200\n",
     "!python -m mathcompose.eval.run processbench --adapter runs/verifier_v --tag tuned --maj-k 8 --limit 200\n",
     "!python -m mathcompose.eval.run prmbench    --adapter runs/verifier_v --tag tuned --limit 300\n",
-    "!python -m mathcompose.eval.run report      # -> results/RESULTS.md"
+    "!python -m mathcompose.eval.run report"
    ]
   },
   {
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "## 6. Push the trained V adapter to the Hub (deliverable #2)"
+    "## 6. Push the trained V adapter (no retrain) \u2014 deliverable #2"
    ]
   },
   {
@@ -101,8 +98,11 @@
    "execution_count": null,
    "outputs": [],
    "source": [
-    "!python -m mathcompose.train.train --task v --config configs/verifier_v.yaml \\\n",
-    "    --push --hub-model-id {USER}/mathcompose-verifier-1.5b"
+    "from huggingface_hub import HfApi, create_repo\n",
+    "MODEL_REPO = f\"{USER}/mathcompose-verifier-1.5b\"\n",
+    "create_repo(MODEL_REPO, exist_ok=True)\n",
+    "HfApi().upload_folder(folder_path=\"runs/verifier_v\", repo_id=MODEL_REPO, repo_type=\"model\")\n",
+    "print(\"pushed ->\", f\"https://huggingface.co/{MODEL_REPO}\")"
    ]
   },
   {
@@ -110,8 +110,8 @@
    "metadata": {},
    "source": [
     "## 7. (Later) Model G + the composition\n",
-    "After V looks good, build G's dataset (teacher=claude-opus-4-8), train it, then run\n",
-    "the V-reranks-G composition \u2014 the headline generator-verifier-gap number."
+    "Build G's dataset (teacher=claude-opus-4-8), train, then run V-reranks-G \u2014 the\n",
+    "headline generator-verifier-gap number."
    ]
   },
   {
@@ -121,7 +121,6 @@
    "outputs": [],
    "source": [
     "# !python -m mathcompose.train.train --task g --config configs/generator_g.yaml\n",
-    "# !python -m mathcompose.eval.run math    --adapter runs/generator_g --tag tuned --limit 200\n",
     "# !python -m mathcompose.eval.run compose --gen-adapter runs/generator_g --ver-adapter runs/verifier_v \\\n",
     "#     --dataset Maxwell-Jia/AIME_2024 --split train --problem-field Problem --answer-field Answer"
    ]