Crownelius/GLM-5.2-CoT-Library

GLM-5.2 — CoT Library

A maintained mirror of publicly-available GLM-5.2 chain-of-thought datasets on Hugging Face — content-verified, deduplicated, and attributed to their original authors.

Dataset Viewer | Parquet

// what this is

A maintained library — a community mirror of publicly-available GLM-5.2 CoT datasets, aggregated, validity-filtered and content-verified, with per-row source attribution in first_source_dataset. It is not Crownelius' own data — every row belongs to the uploader credited below. Reasoning, science and logic-puzzle chain-of-thought plus agentic coding traces.

41,333
verified rows

7
contributing datasets

12
rejected in verification

How a dataset gets in

Every candidate is downloaded and inspected — never judged by its name:

1. Readable — it actually downloads and parses.
2. Structurally valid — rows are QA / chain-of-thought / agent traces with a real question and a substantive response.
3. Right model — content or an in-row model field ties it to GLM-5.2; sets dominated by a rival model are rejected.
4. Not a duplicate — dedup by sha256(row_json) and a payload fingerprint that ignores metadata wrappers.

📜  Citations & attribution — 7 contributing datasets _{(please cite the original authors, not this mirror)}

Rows	Original author & dataset	License
26,407	ianncity — GLM-5.2-Conversation	apache-2.0
6,402	ianncity — GLM-5.2-Logic-Puzzles	apache-2.0
4,308	Jimbock — GLM_5.2_Dataset	apache-2.0
1,960	ianncity — GLM-5.2-Science	apache-2.0
1,744	greghavens — glm-5.2-coding-and-debugging-traces	cc-by-4.0
319	AletheiaResearch — GLM-5.2-Agent	see upstream
193	AletheiaResearch — GLM-5.2-Bench	see upstream

total 41,333 content-verified rows

🚫  Rejected in verification — 12 datasets _{(every one linked so you can check the call)}

Verdict	Dataset	Why
CARD-ONLY	clzoro/GLM-5.1-1000000x	no in-row model field and no model signal in content — provenance is card-asserted only
CARD-ONLY	ianncity/GLM-5.2-Conversation	no in-row model field and no model signal in content — provenance is card-asserted only
CARD-ONLY	hotdogs/uka-glm-5.2	no in-row model field and no model signal in content — provenance is card-asserted only
CARD-ONLY	ianncity/GLM-5.2-Science	no in-row model field and no model signal in content — provenance is card-asserted only
CARD-ONLY	ianncity/GLM-5.2-Logic-Puzzles	no in-row model field and no model signal in content — provenance is card-asserted only
UNVERIFIABLE	mgoin/open-perfectblend-glm5.2-regen	could not be downloaded or parsed (parse_err:ArrowInvalid)
UNVERIFIABLE	ianncity/GLM-5.2-Finance-80000x	could not be downloaded or parsed (info_http=None)
CARD-ONLY	shreyan35/Project-Axiom-1.0-Opus4.7-Kimi2.6-GLM5.2-Deepseek4-Mythos5-Fable5-Qwen3.7	no in-row model field and no model signal in content — provenance is card-asserted only
CARD-ONLY	DavidrPatton/Fable-5-GLM-5.2-Traces	no in-row model field and no model signal in content — provenance is card-asserted only
FAILS GATE	mgoin/GLM-5.2-FP8-magpie-ultrachat	fails the content gate — only 0/50 sampled rows have a real question and a substantive response
FAILS GATE	JessieWei/GLM-5.2-FP8-nemotron-codealpaca	fails the content gate — only 0/50 sampled rows have a real question and a substantive response
CARD-ONLY	Jimbock/GLM_5.2_Dataset	no in-row model field and no model signal in content — provenance is card-asserted only

Provenance & honesty

Every retained row is content-verified as genuine GLM-5.2-family data. That provenance is source-asserted and content-verified — it cannot be cryptographically certified by Zhipu AI. No method exists to prove a block of text came from a specific model. What is guaranteed is narrower and checkable: keyword matches are removed, every row carries a real question and response, cross-model contamination is rejected, and everything excluded is listed above with a reason.

Schema

Column	Type	Meaning
row_hash	string	sha256(row_json) — de-duplication key
first_source_dataset	string	upstream dataset / original author
first_source_config	string	upstream config
first_source_split	string	upstream split
first_source_row_index	int64	index within the upstream source
seen_count	int64	times this canonical row was seen
row_json	string	the full source row as JSON

Loading

from datasets import load_dataset
ds = load_dataset("Crownelius/GLM-5.2-CoT-Library", split="train")

import json
row = json.loads(ds[0]["row_json"])

---

_{Maintained mirror · content-verified · all credit to the original uploaders}