Upload folder using huggingface_hub

data/verifier/README.md

@@ -12,10 +12,11 @@ paragraph-by-paragraph critique ending in `\boxed{{k}}` (`completion`), where `k
 is the 0-based index of the first erroneous step (`-1` = all correct).
 
 - **Labels** are ground truth from **PRM800K** (OpenAI, MIT).
-- **Critiques** distilled from **claude-opus-4-8** (rationalization conditioned on
-  the known label; the boxed index is stamped authoritatively, not trusted to the teacher).
+- **Genuine-detection recipe:** **claude-opus-4-8** critiques each solution *blind*
+  (never shown the answer); a critique is kept only if its predicted index matches
+  the PRM800K gold (rejection sampling, ~77% keep-rate). So the critiques contain
+  real error-finding reasoning, not rationalization.
 - **Deduped** against ProcessBench + MATH-500 (no eval contamination).
-- `train.jsonl`/`val.jsonl` are class-balanced (~50/50 erroneous/all-correct);
-  `*_full.jsonl` keep the natural PRM800K distribution.
+- Naturally ~50/50 erroneous/all-correct (no rebalancing needed).
 
 Built by the mathcompose harness. Eval target: ProcessBench first-error F1.

scripts/publish_hf.py

@@ -33,11 +33,12 @@ paragraph-by-paragraph critique ending in `\\boxed{{k}}` (`completion`), where `
 is the 0-based index of the first erroneous step (`-1` = all correct).
 
 - **Labels** are ground truth from **PRM800K** (OpenAI, MIT).
-- **Critiques** distilled from **claude-opus-4-8** (rationalization conditioned on
-  the known label; the boxed index is stamped authoritatively, not trusted to the teacher).
+- **Genuine-detection recipe:** **claude-opus-4-8** critiques each solution *blind*
+  (never shown the answer); a critique is kept only if its predicted index matches
+  the PRM800K gold (rejection sampling, ~77% keep-rate). So the critiques contain
+  real error-finding reasoning, not rationalization.
 - **Deduped** against ProcessBench + MATH-500 (no eval contamination).
-- `train.jsonl`/`val.jsonl` are class-balanced (~50/50 erroneous/all-correct);
-  `*_full.jsonl` keep the natural PRM800K distribution.
+- Naturally ~50/50 erroneous/all-correct (no rebalancing needed).
 
 Built by the mathcompose harness. Eval target: ProcessBench first-error F1.
 """

src/mathcompose/train/colab_train.ipynb

@@ -82,7 +82,8 @@
    "outputs": [],
    "source": [
     "!python -m mathcompose.eval.run processbench --tag base  --maj-k 1 --limit 100 --splits gsm8k\n",
-    "!python -m mathcompose.eval.run processbench --adapter runs/verifier_v --tag tuned --maj-k 4 --limit 100 --splits gsm8k\n",
+    "# greedy (Maj@1) is the fair metric until V is strong; Maj@k only helps a confident verifier\n",
+    "!python -m mathcompose.eval.run processbench --adapter runs/verifier_v --tag tuned --maj-k 1 --limit 100 --splits gsm8k\n",
     "# full run when the fast read looks right:\n",
     "# !python -m mathcompose.eval.run processbench --tag base  --maj-k 1\n",
     "# !python -m mathcompose.eval.run processbench --adapter runs/verifier_v --tag tuned --maj-k 8\n",