zgcagi/ZGCM-1-Data

A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search

Zhongguancun Academy · Zhongguancun Institute of Artificial Intelligence

📄 Tech Report · 🤗 Model · 🤗 Data · 📊 Results · 💻 Training Code · 💬 WeChat Community

Introduction

ZGCM-1 is a 7.39B-parameter dense language model trained from scratch, built for mathematical reasoning and tool-assisted search. It combines deliberate internal thinking with active information gathering, supporting 256K-token context and both thinking and direct-response modes in a single model.

The project brings together an efficient hybrid-attention architecture, FP8 training with Muon, progressive long-context mid-training, and general-agentic supervised fine-tuning. Researcher-directed AI agents contribute throughout development, from data curation and cluster operations to evaluation and deployment.

ZGCM-1-Data is the training collection behind that model: a bilingual Chinese-English corpus for large language model training, covering general pretraining, capability-oriented and long-context midtraining, and supervised fine-tuning.

The collection includes web pages, books, papers, PDF/OCR text, encyclopedic content, code, mathematics, question answering, reasoning, instruction following, tool use, and software engineering trajectories.

Evaluation Results

ZGCM-1, the 7.39B model trained on this collection, reaches the following scores. Results are from the technical report, using the 256K SFT checkpoint in thinking mode. Non-agentic evaluations use temperature 1.0, top-p 1.0, and mean pass@1 over 32 runs unless otherwise specified.

Selected reasoning benchmarks

Benchmark (%)	ZGCM-1	DeepSeek-R1-0528-Qwen3-8B	MiniCPM4.1-8B	Qwen3-8B	Olmo 3 7B Think
MATH-500	97.13	96.32	95.60	96.20	95.10
AIME 2024	80.62	83.33	83.33	80.00	71.60
AIME 2025	73.33	75.21	73.33	63.33	64.60
AIME 2026	75.00	69.17	71.67	66.67	66.16
HMMT 2025	70.42	61.50	52.50	43.33	43.89
HMMT 2026	59.48	51.52	46.21	45.45	43.94

Selected rows and models from Table 2; bold marks the best score in each displayed row. The full evaluation covers 20 benchmarks, including code, knowledge, and instruction following.

Agentic search

Benchmark	ZGCM-1 (%)	Setting
WebWalkerQA	63.09	Web search and page reading
BrowseComp	19.43	Web search and page reading
GAIA (text-only)	42.52	Web search and page reading
Binary Function Search	62.00	31/50 exact function-entry matches using Ghidra tools

Source: Tables 3–4. Web research allows up to 64 search-and-read steps. Binary Function Search uses a separate protocol on 50 tasks from 10 held-out projects.

Dataset Overview

Stage	Configuration	Size	Primary purpose
Pretrain	zgcm-1-pretrain-stage1	990.1B tokens	Curriculum learning
Pretrain	zgcm-1-pretrain-stage2	3,201.3B tokens	Continued pretraining on the full data mixture
Midtrain	zgcm-1-midtrain-16k	Approx. 180B tokens	Capability-oriented training
Midtrain	zgcm-1-midtrain-64k	Approx. 240B tokens	64K long-context training
Midtrain	zgcm-1-midtrain-256k	Approx. 180.9B tokens	256K long-context training
SFT	zgcm-1-sft	Approx. 4.57M samples; 19.46B packed tokens	Instruction, reasoning, code, tool use, and context lengths up to 256K

Stage	Total tokens
Pretrain	4,191.4B
Midtrain	Approx. 600.9B
SFT	19.46B packed tokens

Data Composition

Pretrain

The pretraining mixture contains approximately 4.2T tokens. Stage 1 supports curriculum learning, while Stage 2 continues pretraining on a broader mixture. The data covers web text, code, academic documents and OCR-processed PDFs, mathematics, LaTeX papers, reasoning, and encyclopedic content.

Major source families include Ultra-FineWeb, FinePDFs, OLMoCR, Stack-Edu, GitHub code, Nemotron, FineMath, OpenWebMath, Proof-Pile, arXiv, and FineWiki, together with other public or appropriately licensed sources.

Midtrain

The midtraining data is organized into nine categories:

Category	Contents
Web	High-quality web pages and translated web content
Knowledge	Books, open educational resources, papers, PDF/OCR text, encyclopedic content, and domain references
Code	Code completion, question answering, rewriting, review, scientific computing, and software development tasks
Math	Mathematical textbooks, question answering, and reasoning data
QA	Large-scale question answering derived from web content
Reasoning	General reasoning, STEM, and research-oriented question answering
Instruction	General instruction-following and conversational data
Agentic	Tool use, environment interaction, and software engineering trajectories
Pretrain replay	Replay of general pretraining data

Web and Knowledge data are further divided into three quality tiers. Long-context configurations combine shorter-prefix data with examples from the target length range to retain general capabilities while increasing long-context coverage.

Length groups are defined as follows:

Length group	Token range
le16k	(0, 16K]
gt16k_le64k	(16K, 64K]
gt64k	>64K, used in the 256K configuration

SFT

The SFT configuration contains approximately 4.57M samples from 50 source-level components. It covers:

• general instruction following and knowledge question answering;
• mathematical, scientific, and code reasoning;
• multilingual and safety-oriented tasks;
• search, tool use, and multi-turn conversations;
• software engineering and agent trajectories; and
• context lengths up to 256K.

Each sample retains a stable source identifier for provenance tracking and analysis.

Detailed source composition, mixture weights, curriculum design, and length distributions are described in the accompanying technical report [1].

Dataset Structure

All six configurations use a shared Parquet schema. Full-text Pretrain and Midtrain records primarily use the text field:

{
  "text": "Model-visible training text.",
  "source": "source-name",
  "source_group": "normalized-source-group",
  "upstream_repos": ["organization/dataset"],
  "upstream_revision": "revision",
  "licenses": ["upstream-license-identifier"],
  "release_mode": "full_text",
  "category": "web",
  "token_count": 1024,
  "metadata_json": "{...}"
}

SFT

{
  "id": "sample-id",
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "tools_json": "[]",
  "source": "source-name",
  "source_group": "normalized-source-group",
  "upstream_repos": ["organization/dataset"],
  "upstream_revision": "revision",
  "licenses": ["upstream-license-identifier"],
  "release_mode": "full_text",
  "metadata_json": "{...}"
}

Variable-structure fields are serialized as JSON strings. tools_json stores tool definitions, locator_json stores source locator fields, and metadata_json stores additional source metadata. For full-text records, licenses records the applicable upstream license at the leaf-source or item level. database_license: odc-by-1.0 identifies a database-level license and does not replace item-level licenses. Public sources without a declared upstream license use license_status: not_declared_upstream.

Data files use Zstandard-compressed Parquet and are stored under data/<config>/. Code derived from GHArchive is not included as full text in the training files. Commit-addressable sample and repository indexes are available under indexes/gharchive-clean-v6-core/.

Index-only Records

For sources that cannot be redistributed directly, require separate acceptance of upstream terms or access approval, or require users to verify licensing with the upstream provider, the dataset provides index_only records. These records contain source locators but no model-visible text:

{
  "id": "sample-id",
  "source": "source-name",
  "upstream_repos": ["organization/dataset"],
  "upstream_revision": "revision",
  "token_count": 1024,
  "locator_json": "{\"upstream_id\":\"upstream-record-id\",\"text_sha256\":\"...\"}",
  "release_mode": "index_only",
  "content_withheld_reason": "upstream_access_required"
}

index_only records do not contain text, content, prompts or responses, messages, reasoning, solutions, tool calls, observations, trajectories, or free-form metadata that may expose source text. Each index uses a fixed upstream revision and provides either a record identifier or a file path and line number that resolves within that revision.

Code indexes may additionally include repo_name, repo_url, path, blob_id, commit identifiers, language, and detected license information, but do not include source code text.

Sources Requiring Separate Upstream Access

The following sources require users to accept upstream terms or request access separately. This dataset provides index records for these sources without redistributing their full text:

Upstream source	Access requirement	Stage
NVIDIA Nemotron-CC v2 and v2.1	Manual Hugging Face approval and acceptance of the NVIDIA Data Agreement	Midtrain
NVIDIA Nemotron-CC-Code v1 and Nemotron-Pretraining-Code v1/v2	Manual Hugging Face approval and acceptance of the NVIDIA Data Agreement	Pretrain and Midtrain
NVIDIA Nemotron-CC-Math v1	Hugging Face access confirmation and acceptance of the NVIDIA Data Agreement	Pretrain and Midtrain
BigCode The Stack v2	Hugging Face access confirmation; individual files remain subject to their repository licenses	Pretrain and Midtrain replay
GAIR daVinci-Dev	Hugging Face access confirmation; the dataset contains multiple upstream licenses	Midtrain
OpenBMB UltraData-SFT-2605	Hugging Face access confirmation	SFT

The following sources are also distributed as indexes only:

Source	Index format
GitGud code records with license=unknown	No repository license was detected upstream
Code records derived from GHArchive	Located by repository, full Git commit SHA, and relative file path; source code remains available from the upstream repository

When an upstream source does not declare a single license or contains multiple licenses, license_status records the corresponding state. Users should consult the upstream repository and revision fields together with the relevant data card and access terms.

Data Processing

Depending on the data type, processing includes the following steps:

1. Text extraction, schema normalization, and source attribution.
2. Removal of invalid Unicode, control characters, empty text, and malformed records.
3. Exact deduplication, with additional conservative near-deduplication for Midtrain.
4. Quality filtering using programmatic rules, source quality scores, quality classifiers, and checks tailored to OCR, code, mathematics, and tool-use trajectories. High-risk erroneous samples are also removed from selected mathematical sources.
5. Validation of tool calls, tool responses, and message ordering in tool-use and agent data.
6. Token counting and length grouping with the GLM-5.1 tokenizer.
7. Sample-level sampling and shuffling by source, category, and length.

To improve coverage of scarce capabilities, a small number of carefully filtered, capacity-limited sources are upsampled in a controlled manner. Source identifiers are retained for all samples, and repeat-sampling information is stored in metadata_json where available.

The accompanying technical report provides additional details on processing, parameters, and deduplication [1].

Loading

from datasets import load_dataset

dataset = load_dataset(
    "zgcagi/ZGCM-1-Data",
    name="zgcm-1-midtrain-64k",
    split="train",
    streaming=True,
)

See the repository README.md and file list for the available configurations.

Intended Use

This dataset is intended for:

• general language model pretraining;
• capability development in code, mathematics, reasoning, question answering, and knowledge;
• long-context training from 16K to 256K tokens;
• supervised fine-tuning for instruction following, tool use, and software engineering; and
• research on data mixtures, quality stratification, and training-data methodology.

Limitations

• The collection primarily covers Chinese, English, and common programming languages; representation is not uniform across languages.
• Accuracy, formatting consistency, and reasoning reliability may vary across sources.
• Quality filtering and deduplication improve aggregate data quality but do not replace task-specific downstream evaluation.
• The coverage and writing style of synthetic data may reflect the models used to generate it.
• Additional domain- and application-specific evaluation is recommended for specialized or high-impact use cases.

License

This collection combines data under multiple upstream licenses and is therefore labeled other at the Hugging Face repository level.

• Each sub-source or item remains subject to its upstream license and terms. Provenance, revision, and license information is provided in upstream_repos, upstream_revision, licenses, and license_status.
• database_license: odc-by-1.0 denotes a database-level license and does not replace rights or restrictions attached to individual documents or code files.
• Records under CC-BY-NC-4.0 may only be used as permitted by that license, including its non-commercial restriction.
• index_only records contain no source text; an index does not grant or replace an upstream license.
• Original processing scripts, manifest templates, and documentation that the project is authorized to license are released under the MIT License. The MIT License does not cover third-party data.

Users are responsible for reviewing the record-level metadata, upstream data cards, and applicable terms before use.

WeChat Community

Scan the QR code to join the ZGCM-1 community group. Click the image to open it at full size. If the code has expired, please open a Discussion and ask the maintainers for the latest one.

References

[1] ZGCM Team. ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search. 2026. PDF