CDMT data and raw evaluation results
Companion data for the models wenyii/cdmt-qwen2.5-3b and wenyii/cdmt-checkpoints-extra.
| Folder | Contents |
|---|---|
data/factorial10k_unique/ |
Math (DeepMath-103K) and Search (NQ + HotpotQA) 10,000-prompt pools, near-duplicate exclusions, nested 1k pools, SFT traces (*_sft_train10000.jsonl) and the per-seed SFT parquets (*_sft_exposures80000_s{seed}.parquet), RL prompt parquets (*_train10000.parquet). |
data/factorial10k_code_io_v2/ |
Repaired Code pool (TACO + DeepCoder-Preview; reference-solution traces), SFT and RL parquets. |
data/factorial10k_mix_code_io_v2/ |
Mix (1/3 each source) SFT, GRPO, and OPD parquets. |
data/math_full_val.parquet, data/search_full_val.parquet, data/clean_code_io_v2_full_val.parquet |
Evaluation panels (200 items used). |
data/code_debug_training_v4_frozen/, data/code_revision_v1/ |
CodeDebug fixed-opportunity panel (199 items) and Code-debug training data. |
data/nonce_private/ |
Synthetic private-fact (nonce) corpus and panels (generator seed 20260730). |
data/bench_clean_v2/ |
Held-out external panels (Bamboogle, 2Wiki, MuSiQue, MATH500, GSM8K, AIME 2024, HumanEval, MBPP). LiveCodeBench is not redistributed; rebuild it with the builder in the code repository. |
data/clean_sft_search_*, data/sft_clean_search_xteacher_* |
Teacher-choice and family-experiment Search SFT data. |
manifests/ |
Build and retriever manifests with sha256 of every file. |
retriever/e5_IVF.index |
The FAISS IVF index actually served (built from the upstream Search-R1 e5_Flat.index with unseeded clustering, so it is shipped rather than rebuilt). Corpus wiki-18.jsonl and encoder intfloat/e5-base-v2 come from upstream (see the retriever manifest). |
eval_results/<panel>/<model_tag>/<tag>__<target>__<arm>__d0.json |
Raw evaluation outputs with full trajectories (messages, actions) and metrics, including the checkpoint path loaded. |
data_manifest.json |
Historical manifest subset: path, bytes, SHA-256, rows, role, builder and license. The complete 196-file inventory is maintained in the GitHub data guide, including teacher shards and files listed in the recovery supplement. |
Licenses: DeepMath-103K MIT; TACO Apache-2.0; DeepCoder-Preview MIT; NQ CC-BY-SA-3.0; HotpotQA CC-BY-SA-4.0 (share-alike applies to Search and Mix files); Wikipedia CC-BY-SA-3.0. SFT traces for Math and Search were generated with Qwen2.5-72B-Instruct (Built with Qwen). Code SFT traces are built from dataset reference solutions.
Historical training inputs recovered from HPC (2026-11-20)
manifests/DATA_RECOVERY_20261120.tsv lists 41 additional files by exact name, row count and SHA-256:
the original Math/Code teacher JSONL and aligned RL/Code-IO datasets, the early Math/Code/Search
matched SFT train/validation Parquets, the Gemma/Llama Search-SFT train/validation Parquets, and six rendered files for the legacy Search-SFT seeds.
The earlier data_manifest.json still describes the original release. Find the per-method and
per-benchmark instructions in docs/DATA_INVENTORY.md in the authorized code repository.
These files retain the upstream attribution and terms described in the data source guide:
DeepMath-103K (MIT), TACO (Apache-2.0), DeepCoder-Preview (MIT), FlashRAG/HotpotQA (CC BY-SA),
and Qwen2.5 teacher outputs.
provenance/codedebug-candidate-reproduction-20261120.tar.gz contains the archived five-seed
CodeDebug candidate and its frozen evidence, together with checksums. The candidate is separate
from the submitted paper. Its CPU audit recomputes saved statistics without running model-generated code.
The complete derived LiveCodeBench input remains outside this archive; use the pinned upstream
source and the builder recorded in the code documentation.
- Downloads last month
- 100