Upload verifier reward-model dataset and provenance report

README.md

@@ -0,0 +1,74 @@
+# Instruction-conditioned verifier reward-model dataset
+
+This folder contains a bounded, reproducible mixture for a Qwen3 4B LoRA reward model. Each row is a single candidate to score. The model is trained to emit `YES` for a candidate that satisfies the row's `custom_instruction` and `NO` otherwise; downstream users can use the Yes/No logit margin as the continuous reward.
+
+The dataset is intended to remain private. The model weights can be released separately. The build script streams source datasets and stores source IDs, provenance, evidence regime, and supervision type in every row.
+
+## Build locally
+
+The current environment already has `datasets`, `huggingface_hub`, and `pyarrow`. Otherwise:
+
+```powershell
+python -m pip install -r requirements.txt
+python build_dataset.py
+```
+
+The default build is intentionally bounded at roughly 20k candidate rows. Adjust quotas if desired:
+
+```powershell
+python build_dataset.py --math-examples 12000 --ultramix-pairs 2500 --max-total 30000
+```
+
+Output files are written under `data/` as Parquet shards. Build provenance and counts are written to `reports/build_manifest.json` and `reports/dataset_stats.json`.
+
+Validate the generated shards and group-level split isolation:
+
+```powershell
+python validate_dataset.py
+```
+
+## Upload as a private Hub dataset
+
+Create a user token with write access, then set it only in the current PowerShell process:
+
+```powershell
+$env:HF_TOKEN = "hf_..."
+python push_to_hub.py --repo-id YOUR_USERNAME/verifier-rm-data
+```
+
+The uploader creates the dataset repository as private by default and uploads only the data, README, and JSON reports. It never writes the token to disk.
+
+## Normalized schema
+
+The core fields are:
+
+- `custom_instruction`: the criterion that defines what YES means.
+- `current_completion`: the user prompt or preceding conversation.
+- `text_to_score`: the candidate solution/response/trace.
+- `reference_solution`: optional authoritative solution.
+- `privileged_context`: optional authoritative hint for v1.
+- `evidence_mode`: `none`, `reference`, `privileged`, or `both`.
+- `target_label` / `target_binary`: `YES`/`NO` supervision target.
+- `supervision_type`: `absolute_verifiable` or `pairwise_preference`.
+- `label_provenance` and `label_confidence`: source and trust metadata.
+- `group_id`: keeps candidates from one underlying problem in the same split.
+
+Splits are assigned by `group_id`, not by row, so paired candidates and math generations cannot leak across train/validation/test.
+
+## Included source families
+
+- `open-r1/OpenR1-Math-220k`: math reasoning traces with Math-Verify correctness labels. The builder creates several evidence views while retaining the verified label.
+- `aladinDJ/ultramix-DPO-annotated`: quality-filtered, task-balanced preference pairs with math, reasoning, coding, and general tasks.
+- `nvidia/HelpSteer3` (`preference`): human-annotated multi-turn preferences across general, STEM, code, and multilingual domains.
+- `coseal/CodeUltraFeedback_binarized`: coding preference pairs with ratings and model provenance.
+- `Skywork/Skywork-Reward-Preference-80K-v0.2`: general reward-preference pairs.
+- `R-I-S-E/RISE-Judge-SFT-20K`: position-verified judge comparisons derived from math and preference sources, normalized into candidate-level YES/NO rows.
+- `iknow-lab/JudgeBias-DPO-RefFree` (`validation_1000.parquet` only): a bounded anti-shortcut/materials-science slice designed to preserve semantic equivalence while detecting injected errors. The full upstream repository has inconsistent parquet schemas, so the builder reads only this stable small file directly.
+
+Preference pairs remain marked as pairwise rather than being presented as absolute truth. That distinction matters because a chosen response can be better than a rejected response while still being imperfect.
+
+## Deliberate exclusions
+
+PRM800K is not used as the main math source because this first release emphasizes richer full-solution and preference variation. Evaluation-only sets and sources with inconsistent Hub schemas are not forced into the build; the manifest records any source that fails to stream.
+
+Check each upstream dataset's current license and terms before redistributing derived data. This local/private build does not grant redistribution rights.

SOURCE_REVIEW.md

@@ -0,0 +1,38 @@
+# Source review and mixture rationale
+
+This is the source-selection record for the local v1 build. It is deliberately separate from the normalized data so the model artifact can be released without publishing the derived training set.
+
+## Included
+
+| Source | Role in v1 | Important caveat |
+| --- | --- | --- |
+| [open-r1/OpenR1-Math-220k](https://huggingface.co/datasets/open-r1/OpenR1-Math-220k) | Main math source; multiple reasoning traces and Math-Verify correctness metadata | The builder uses only a bounded streamed subset and creates evidence views; it does not claim every trace is human-verified. |
+| [aladinDJ/ultramix-DPO-annotated](https://huggingface.co/datasets/aladinDJ/ultramix-DPO-annotated) | Balanced general, reasoning, math, coding and instruction-following preferences with curation metadata | Preference pairs are kept as pairwise supervision, not treated as absolute truth. |
+| [nvidia/HelpSteer3](https://huggingface.co/datasets/nvidia/HelpSteer3) | Human-annotated multi-turn preference coverage across general, STEM, code and multilingual domains | Overall preference has a three-point direction scale; ties are skipped. |
+| [coseal/CodeUltraFeedback_binarized](https://huggingface.co/datasets/coseal/CodeUltraFeedback_binarized) | Code-specific preference and rating signal | Identical chosen/rejected pairs are skipped. |
+| [Skywork/Skywork-Reward-Preference-80K-v0.2](https://huggingface.co/datasets/Skywork/Skywork-Reward-Preference-80K-v0.2) | Additional general reward-preference diversity | The source has limited task metadata, so rows use the generic rubric and preserve the upstream source tag. |
+| [R-I-S-E/RISE-Judge-SFT-20K](https://huggingface.co/datasets/R-I-S-E/RISE-Judge-SFT-20K) | Judge-style comparisons with a generated rationale and position-swap verification | The builder extracts the user question and A/B candidates, discarding the long judge rationale from the scored candidate. |
+| [iknow-lab/JudgeBias-DPO-RefFree](https://huggingface.co/datasets/iknow-lab/JudgeBias-DPO-RefFree) | Bounded anti-shortcut/materials-science pairs for representation invariance and injected-error detection | The builder uses only the stable `validation_1000.parquet` file because the full upstream parquet set currently has inconsistent schemas. |
+
+## Explored but not in the default build
+
+| Source | Decision |
+| --- | --- |
+| [OpenAI PRM800K](https://github.com/openai/prm800k) | Deliberately not used as the main math source because v1 is intended to be richer than simple step labels. |
+| Math-Shepherd | Useful process-supervision reference, but not included in the default mixture; its automated step labels would add another label regime before the core schema is stable. |
+| Full JudgeBias train/validation parquet set | Not used because the Hub parquet files currently expose inconsistent columns and fail the streaming cast check; only the stable 999-row `validation_1000.parquet` slice is included. |
+| [amazon/CodePrefBench](https://huggingface.co/datasets/amazon/CodePrefBench) | Kept as an evaluation candidate rather than training data; its Hub card identifies a non-commercial license and an evaluation/test orientation. |
+| [cracklinoatbran/reward_hacking_policy_1073](https://huggingface.co/datasets/cracklinoatbran/reward_hacking_policy_1073) | Prompt-only reward-hacking evaluation set. It needs fresh candidate generations and verified hack/legitimate labels, so it is reserved for a later anti-shortcut generation stage. |
+| [hlyn-labs/prompt-injection-judge-dataset-v1](https://huggingface.co/datasets/hlyn-labs/prompt-injection-judge-dataset-v1) | Security-judge conversations are not directly compatible with the v1 candidate-level correctness contract; reserve for a later adversarial/instruction-injection slice. |
+
+## Mixture properties of the generated artifact
+
+The current build contains 20,799 unique candidate rows in three group-isolated Parquet splits:
+
+- 8,000 absolute math-verification rows.
+- 12,799 pairwise preference/judge rows across general, STEM, code, materials science, and math/judgment sources.
+- 14,816 rows with no evidence, 1,972 with a reference solution, 2,022 with privileged context, and 1,989 with both.
+- 18,712 train, 1,094 validation, and 993 test rows.
+- 20,799 unique `example_id` values and 10,265 unique `group_id` values.
+
+The source datasets are not all licensed identically. This folder is therefore a private local artifact by default; check each upstream dataset's current license and terms before redistributing derived data. The model-weight release should document the source mixture without publishing these Parquet files unless redistribution is separately cleared.

data/test-00000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:db57f0ba03ca862864a98f8c518be6c569f57e505a8f3cef955d76716da28960
+size 3693159

data/train-00000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8901a51c036fcc9ba15b68d461a113c133407f2c4cba295872296571a5691d1d
+size 69366026

data/validation-00000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a583db42940bfe05c50b6db0aceca8ccf1d6e98df36d395b7ad0e2cdfa862622
+size 3989225

reports/build_manifest.json

@@ -0,0 +1,121 @@
+{
+  "schema_version": "verifier-rm-1.0",
+  "quotas": {
+    "output": "C:\\Users\\sunco\\OneDrive\\Desktop\\agentic-rl\\verifier-dataset",
+    "math_examples": 8000,
+    "ultramix_pairs": 1500,
+    "helpsteer3_pairs": 1200,
+    "code_pairs": 800,
+    "skywork_pairs": 1000,
+    "rise_pairs": 1600,
+    "judgebias_pairs": 400,
+    "max_total": 24000
+  },
+  "sources": {
+    "openr1_math": {
+      "status": "ok",
+      "scanned": 3876,
+      "accepted": 8000,
+      "skipped": 0
+    },
+    "ultramix": {
+      "status": "ok",
+      "scanned": 1500,
+      "pairs": 1500,
+      "accepted": 3000,
+      "skipped": 0
+    },
+    "helpsteer3": {
+      "status": "ok",
+      "scanned": 1216,
+      "pairs": 1200,
+      "accepted": 2400,
+      "skipped": 16
+    },
+    "code_ultrafeedback": {
+      "status": "ok",
+      "scanned": 806,
+      "pairs": 800,
+      "accepted": 1600,
+      "skipped": 6
+    },
+    "skywork": {
+      "status": "ok",
+      "scanned": 1000,
+      "pairs": 1000,
+      "accepted": 2000,
+      "skipped": 0
+    },
+    "rise_judge": {
+      "status": "ok",
+      "scanned": 3316,
+      "pairs": 1600,
+      "accepted": 3198,
+      "skipped": 1716
+    },
+    "judgebias": {
+      "status": "ok",
+      "scanned": 999,
+      "pairs": 400,
+      "accepted": 800,
+      "skipped": 0
+    }
+  },
+  "deduplication": {
+    "before": 20998,
+    "after": 20799
+  },
+  "final_stats": {
+    "total_examples": 20799,
+    "splits": {
+      "train": 18712,
+      "validation": 1094,
+      "test": 993
+    },
+    "sources": {
+      "open-r1/OpenR1-Math-220k": 8000,
+      "aladinDJ/ultramix-DPO-annotated": 3000,
+      "nvidia/HelpSteer3": 2255,
+      "coseal/CodeUltraFeedback_binarized": 1600,
+      "Skywork/Skywork-Reward-Preference-80K-v0.2": 2000,
+      "R-I-S-E/RISE-Judge-SFT-20K": 3150,
+      "iknow-lab/JudgeBias-DPO-RefFree": 794
+    },
+    "domains": {
+      "math": 8142,
+      "code": 4485,
+      "reasoning": 456,
+      "Information seeking": 1194,
+      "Advice seeking": 76,
+      "Planning": 106,
+      "Data analysis": 162,
+      "Creative writing": 144,
+      "Brainstorming": 66,
+      "Editing": 22,
+      "Role playing": 2,
+      "general": 2000,
+      "math_or_general_judgment": 3150,
+      "materials_science": 794
+    },
+    "evidence_modes": {
+      "privileged": 2022,
+      "none": 14816,
+      "both": 1989,
+      "reference": 1972
+    },
+    "supervision_types": {
+      "absolute_verifiable": 8000,
+      "pairwise_preference": 12799
+    },
+    "labels": {
+      "YES": 11493,
+      "NO": 9306
+    },
+    "candidate_spans": {
+      "whole_completion": 20799
+    },
+    "avg_candidate_chars": 7476.3,
+    "max_candidate_chars": 72219,
+    "unique_groups": 10265
+  }
+}

reports/dataset_stats.json

@@ -0,0 +1,53 @@
+{
+  "total_examples": 20799,
+  "splits": {
+    "train": 18712,
+    "validation": 1094,
+    "test": 993
+  },
+  "sources": {
+    "open-r1/OpenR1-Math-220k": 8000,
+    "aladinDJ/ultramix-DPO-annotated": 3000,
+    "nvidia/HelpSteer3": 2255,
+    "coseal/CodeUltraFeedback_binarized": 1600,
+    "Skywork/Skywork-Reward-Preference-80K-v0.2": 2000,
+    "R-I-S-E/RISE-Judge-SFT-20K": 3150,
+    "iknow-lab/JudgeBias-DPO-RefFree": 794
+  },
+  "domains": {
+    "math": 8142,
+    "code": 4485,
+    "reasoning": 456,
+    "Information seeking": 1194,
+    "Advice seeking": 76,
+    "Planning": 106,
+    "Data analysis": 162,
+    "Creative writing": 144,
+    "Brainstorming": 66,
+    "Editing": 22,
+    "Role playing": 2,
+    "general": 2000,
+    "math_or_general_judgment": 3150,
+    "materials_science": 794
+  },
+  "evidence_modes": {
+    "privileged": 2022,
+    "none": 14816,
+    "both": 1989,
+    "reference": 1972
+  },
+  "supervision_types": {
+    "absolute_verifiable": 8000,
+    "pairwise_preference": 12799
+  },
+  "labels": {
+    "YES": 11493,
+    "NO": 9306
+  },
+  "candidate_spans": {
+    "whole_completion": 20799
+  },
+  "avg_candidate_chars": 7476.3,
+  "max_candidate_chars": 72219,
+  "unique_groups": 10265
+}