PROVENANCE.md

Provenance and Terms

Release identity

• Hugging Face repository: r0b0tlab/qwen3.8-max-glm5.2-distillation-51389
• Canonical release size: 51,389 rows
• Split counts: 46,250 train, 2,569 validation, 2,570 test
• Input JSONL SHA-256: 0733182e2ae59d357e5022c4971ad0cae2482fd1762f8f1447c3ae2c32c517b9
• Materialization report SHA-256: a32836e98511d4902c666dc8ef1d8fc4da12c15fae14689efc54536db7667648
• Native renderer used for the GLM view: zai-org/GLM-4.7-Flash@7dd20894a642a0aa287e9827cb1a1f7f91386b67

The public Parquet shards are copied byte-for-byte from data/sota_merge/release. The release report records each shard's expected byte count and SHA-256; manifest.json records the corresponding package paths and hashes.

Composition

The selected release contains 46,082 admitted Qwen-derived rows and 5,307 admitted GLM-derived rows. The Qwen input was the immutable 49,772-row base selection after repair, reference checks, deduplication, and admission. The GLM input was a 15,000-row deterministic tool-trajectory generation run; the public 5,307 rows are the rows admitted after exact replay, structural checks, degeneration checks, and independent family terminal-state oracles.

The source-family counts below are copied from the deterministic selection report:

Source family	Selected rows
Evol-Code	8,098
MetaMathQA	6,267
CodeAlpaca	5,586
tulu-3	5,395
glm5.2-agent-tool-synthetic	5,307
SciQ	3,947
NuminaMath-CoT	2,870
CommonsenseQA	2,816
OrcaMath	2,319
Dolly	1,710
QASC	1,682
ARC-Easy	971
gsm8k	640
MATH/algebra	578
OpenBookQA	574
ARC-Challenge	473
MATH/intermediate_algebra	423
MATH/prealgebra	393
MATH/number_theory	239
MATH/precalculus	214
MATH/counting_and_probability	213
MATH/geometry	196
MBPP	183
IFEval	168
HumanEval	127

Pinned source references

• Base Hugging Face release: r0b0tlab/qwen3.8-max-distillation-50k, revision ab9f8b289423c249fc0054507f045a12efb54b1b.
• GLM source run: full-20260724-175443, source JSONL SHA-256 71f34bba1cb29a831fd1f26b51f0b787d9d4ad2bc51c93cb1af5084759e128e1.
• GLM immutable tool schema SHA-256: 0a8235542f3e86cb3508ca76bbdb79c94ef0c6569e0752f039bec581f0848960.
• Upstream reference snapshots used during admission include pinned NuminaMath-CoT and MetaMathQA revisions recorded in the source reports.

The canonical view retains source_repository, source_revision, source_license, source_split, and source_item_id. Those fields are provenance metadata, not a substitute for reviewing the current upstream license or terms.

Source rights and terms

This is a mixed-source release and is intentionally labeled other, not a blanket permissive license. Known upstream rights include CC BY-NC-SA 4.0 for Evol-Instruct-Code, CC BY-NC 3.0 for SciQ, CC BY-SA 3.0 for Dolly, CC BY 4.0 for MBPP and QASC, Apache 2.0 for IFEval and NuminaMath-CoT, MIT for GSM8K and MetaMathQA, and ODC-BY with mixture-specific considerations for Tulu 3. Some source rows have unresolved or source-specific terms; inspect the per-row metadata and upstream cards.

The teacher-generated rows also carry service-term risk. Alibaba Cloud Model Studio International Product Terms, section 4.48(d)(v), restrict use of Model Studio and its outputs to train or develop products or services competing with Alibaba Cloud or its affiliates unless expressly authorized. The source audit also records a noncommercial-research scope and public/commercial redistribution hold pending rights review.

Official terms reference:

https://www.alibabacloud.com/help/en/legal/latest/alibaba-cloud-international-website-product-terms-of-service-v-3-8-0#d48a5c007bamp

These notices are informational and are not legal advice. The intended scope is controlled, noncommercial research only until all applicable rights holders and service terms have been reviewed.

Development holdouts and contamination

The validation and test splits are development holdouts formed from the selected release. They are not independent capability benchmarks. The source mixture includes benchmark-derived items from GSM8K, HumanEval, MBPP, ARC, MATH, IFEval, SciQ, QASC, OpenBookQA, and related sources. Consequently, the development holdouts and overlapping source benchmarks can be contaminated for evaluation. Use fresh, non-overlapping, date-bounded, or privately held tasks for claims.