Document MTMR base origins, preparation and schema

README.md

@@ -106,3 +106,48 @@ configs:
   - split: train
     path: base/train-*
 ---
+
+
+<!-- mtmr-base:start -->
+## MTMR (Multi-Task Multi-Reward) base
+
+`base/train` is the shared, unbalanced prompt pool for model-specific sampling and calibration.
+
+### Sources
+
+- [agentica-org/DeepScaleR-Preview-Dataset](https://huggingface.co/datasets/agentica-org/DeepScaleR-Preview-Dataset), revision `b6ae8c60f5c1f2b594e2140b91c49c9ad0949e29`.
+- [open-r1/DAPO-Math-17k-Processed](https://huggingface.co/datasets/open-r1/DAPO-Math-17k-Processed), revision `31dd309567e3da778038cc87d868b6097a3ccf68`.
+- [PRIME-RL/Eurus-2-RL-Data](https://huggingface.co/datasets/PRIME-RL/Eurus-2-RL-Data), revision `9776b13264b5aaa0b16495fcf086a0a8d86fd655`.
+- [allenai/Dolci-RL-Zero-Code-7B](https://huggingface.co/datasets/allenai/Dolci-RL-Zero-Code-7B), revision `054c0a1f9c5fd52566ad9124aa2e702924803cd3`.
+- [open-r1/verifiable-coding-problems-python_decontaminated-tested](https://huggingface.co/datasets/open-r1/verifiable-coding-problems-python_decontaminated-tested), revision `77ace849f7cc8f10d23945c748f9ff464427f83b`.
+- [sojuL/RubricHub_v1](https://huggingface.co/datasets/sojuL/RubricHub_v1), revision `3837d55971473a872e84879c88f708b8da3ec2ef`.
+Math uses DeepScaleR and DAPO-Math-17k-Processed (`all/train`, English and Chinese).
+Code uses Eurus, Dolci and Open-R1's tested Python problems; IF/Writing use RubricHub's
+`rurbichub_v1_Instruction_Following` and `rurbichub_v1_Writing` files (not `rurbichub_v1_Chat`).
+Rows retain source repository, revision and subset.
+
+### Preparation
+
+- Validate prompts, Math gold parsing, rubric weights and executable IF rules.
+- Remove provably contradictory IF rules and explicit label-only Writing tasks. Deduplicate identical
+  Writing criteria with equal weights; reject conflicting duplicate weights. Ambiguous context/NAME
+  placeholders are flagged for review, not filtered. Cleaning reports are saved locally in `cleaning/`.
+- Normalize Code interfaces/tests; check available Dolci and required Open-R1 reference programs
+  with the training executor (5s per-test timeout, 120s total execution budget).
+  Keep at most 20 uniformly sampled, prompt-seeded tests per question.
+- Render every Code prompt as `### Question / ### Format` from the bare statement (chat, Eurus and
+  Open-R1 wrappers stripped for all sources); drop questions with fewer than 15 words
+  before their first input/example block.
+- Deduplicate normalized prompts across all four tasks; remove exact evaluation-question matches
+  and embedded copies of evaluation questions >=80 characters. Semantic paraphrases are not detected.
+- Require the complete chat prompt to be <=2048 tokens under all three tokenizers:
+  RLLab/olmo-3-1025-7b, Qwen/Qwen3-4B-Base, Qwen/Qwen3.5-2B-Base.
+- Store outer reward weights separately from rubric-criterion weights; length/RM references are
+  filled by the model-specific sampler.
+
+Overlap sources (upstream, not local evaluation files): [Eurus](https://huggingface.co/datasets/PRIME-RL/Eurus-2-RL-Data), [eval-set/aime24](https://huggingface.co/datasets/RLLab/eval-set), [eval-set/aime25](https://huggingface.co/datasets/RLLab/eval-set), [eval-set/amc23](https://huggingface.co/datasets/RLLab/eval-set), [eval-set/minervamath](https://huggingface.co/datasets/RLLab/eval-set), [eval-set/olympiadbench](https://huggingface.co/datasets/RLLab/eval-set), [eval-set/math500](https://huggingface.co/datasets/RLLab/eval-set), [code_generation_lite/v6](https://huggingface.co/datasets/lighteval/code_generation_lite), [MMLU-Pro/default](https://huggingface.co/datasets/TIGER-Lab/MMLU-Pro), [ZebraLogic/grid_mode](https://huggingface.co/datasets/WildEval/ZebraLogic), [humanevalplus](https://github.com/evalplus/humanevalplus_release/releases/download/v0.1.10/HumanEvalPlus.jsonl.gz), [mbppplus](https://github.com/evalplus/mbppplus_release/releases/download/v0.2.0/MbppPlus.jsonl.gz), [IFEval](https://raw.githubusercontent.com/google-research/google-research/master/instruction_following_eval/data/input_data.jsonl), [IFBench](https://raw.githubusercontent.com/allenai/IFBench/1c40f0c10d9b5c5c2f10a175a28007ebb64f7f4d/data/IFBench_test.jsonl), [WritingBench](https://raw.githubusercontent.com/X-PLUG/WritingBench/ae2d5176449b7b769815482641d35926f26793eb/benchmark_query/benchmark_all.jsonl), [WritingBench-validation](https://raw.githubusercontent.com/X-PLUG/WritingBench/7a30aa4bd4305cc912d07b37a491b99097fc5078/benchmark_query/benchmark_all.jsonl), [Arena-Hard-v2](https://raw.githubusercontent.com/lmarena/arena-hard-auto/196f6b826783b3da7310e361a805fa36f0be83f3/data/arena-hard-v2.0/question.jsonl), [GPQA-Diamond](https://openaipublic.blob.core.windows.net/simple-evals/gpqa_diamond.csv), [ResearchQA](https://huggingface.co/datasets/realliyifei/ResearchQA/resolve/main/test.json), [Creative-Writing-v3](https://raw.githubusercontent.com/EQ-bench/creative-writing-bench/c7c3ceef54c40a8ae02dc1c2e1a5e40970fe5c0b/data/creative_writing_prompts_v3.json).
+
+Task counts: `{"code": 35642, "if": 93340, "math": 52981, "writing": 15869}`.
+Refer to the source repositories for dataset licenses.
+
+<!-- mtmr-base:end -->