GLM-5.1-Reasoning-1M-Think-Embeddings (main subset, partial)

Embeddings of the think-block content of every record in the main subset of Jackrong/GLM-5.1-Reasoning-1M-Cleaned, embedded with Qwen/Qwen3-Embedding-0.6B via vLLM.

What this dataset is

Each row is the vector representation of just the reasoning trace (the text between <think>...</think>), not the user prompt and not the final answer. Useful for:

• searching / clustering the original reasoning traces by meaning,
• building a reasoning-style retrieval index over GLM-5.1's own traces,
• distillation / RAG experiments where you want the model's own chain-of-thought in vector form.

Scope of this release

Config	Rows in source	Rows in this release	Status
main	527,737	214,784	partial (filtered by token budget)
Math	22,097	—	not yet released
PHD-Science	103,706	—	not yet released
Multilingual-STEM	92,781	—	not yet released

Total in this release: 214,784 rows (all domain = "main").

The source dataset's main config has 527,737 records; after dropping rows whose extracted think content exceeds 32 000 tokens at the Qwen3 tokenizer (~8 000 tokens ceiling), the kept subset is smaller. Only ~215 k of those filtered rows are included here; the remaining main rows, plus the three other subsets, will be added in subsequent releases.

Source

The cleaned GLM-5.1 reasoning release (Jackrong/GLM-5.1-Reasoning-1M-Cleaned) contains 746,321 records across four subsets. Each record has an output of the form <think>\n{reasoning trace}\n</think>\n\n{final answer}.

Pipeline

1. Extract — regex over each output, joining all <think>...</think> blocks.
2. Filter — skip empty think blocks (≈0 rows) and rows whose extracted text exceeds 32 000 tokens at the Qwen3 tokenizer (≈8 000 tokens ceiling).
3. Embed — POST /v1/embeddings against a vLLM server serving Qwen/Qwen3-Embedding-0.6B (1024-dim, fp32), batched 4 rows per HTTP call, truncate_prompt_tokens=8192 as a safety belt.
4. Persist — shards of 50 000 rows, parquet (snappy), 1 024-dim float32 vectors.

Schema

Column	Type	Description
id	string	id from source (md5 of domain-input-reasoning-answer)
domain	string	always main in this release
input	string	the user prompt from the source record
think_content	string	extracted think-block text, ≤ 32 K chars
think_len_chars	int32	length of think_content in characters
think_len_tokens	int32	length of think_content in Qwen3 tokens
embedding	list	1024-dim dense vector

Load

from datasets import load_dataset
ds = load_dataset("shanaka95/GLM-5.1-Reasoning-1M-Think-Embeddings", split="train")
row = ds[0]
print(row["id"], row["domain"], len(row["embedding"]))  # 1024

Streaming is supported (streaming=True) for cheaper access to a few rows.

Provenance

• Teacher model (the model that produced the reasoning traces): GLM-5.1
• Embedder (this dataset's vectors): Qwen3-Embedding-0.6B
• Upstream: Jackrong/GLM-5.1-Reasoning-1M-Cleaned (apache-2.0)
• Original raw release: Kassadin88/GLM-5.1-1000000x

License

Apache-2.0, inherited from the upstream dataset.