OpenMed onboarding for users in China¶

简体中文版

This guide keeps the first OpenMed install and model download practical on networks where the default PyPI and Hugging Face endpoints are slow or unreachable. It covers package mirrors, a Hugging Face model mirror, a transferable local cache, and the bundled PIPL-oriented policy profile.

The mirrors below are independent third-party services, not OpenMed-operated or endorsed infrastructure. Check your organization's supply-chain policy, pin package and model versions, and validate downloaded artifacts before production use.

For a complete synthetic clinical-note walkthrough, run the Simplified Chinese example or follow the Chinese and Hindi notebook tour.

Install OpenMed from a PyPI mirror¶

For a one-time install through the Tsinghua TUNA mirror:

python -m pip install -i https://pypi.tuna.tsinghua.edu.cn/simple openmed

Install the optional Hugging Face dependencies when this machine will download or inspect model snapshots:

python -m pip install -i https://pypi.tuna.tsinghua.edu.cn/simple "openmed[hf]"

The simple path segment is required. TUNA's PyPI mirror instructions also document the equivalent long service URL.

As an alternative, the Aliyun public mirror supports the same pip interface:

python -m pip install -i https://mirrors.aliyun.com/pypi/simple/ openmed

See the Aliyun PyPI mirror page for its public and ECS-specific endpoints.

Pin the mirror in pip.conf¶

To make TUNA the default for the current user, let pip update the user-level configuration rather than editing an unknown system file:

python -m pip config --user set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
python -m pip config --user get global.index-url

Run python -m pip config debug to see the active pip.conf location. The equivalent file content is:

[global]
index-url = https://pypi.tuna.tsinghua.edu.cn/simple

Replace the URL with https://mirrors.aliyun.com/pypi/simple/ to make Aliyun the default instead. A mirror can lag the canonical index, so pin exact package versions and investigate unexpected version differences before falling back to another index.

Download models through HF-Mirror¶

Install openmed[hf], then set the mirror endpoint before importing huggingface_hub, Transformers, or OpenMed in that process:

export HF_ENDPOINT=https://hf-mirror.com

$env:HF_ENDPOINT = "https://hf-mirror.com"

This is the environment-variable method documented by HF-Mirror. It changes where compatible Hugging Face clients fetch model files; it does not route OpenMed inference to a hosted API.

Pre-download once, then run from the local cache¶

The standard Hugging Face cache lives under ~/.cache/huggingface unless HF_HOME or HF_HUB_CACHE selects another location. Use a dedicated HF_HOME when the cache must be archived or copied to an isolated machine.

1. Warm the cache on a connected machine¶

From an OpenMed source checkout, the example downloads only after an explicit opt-in:

export HF_HOME="$PWD/openmed-hf-cache"
OPENMED_EXAMPLE_ALLOW_DOWNLOAD=1 \
  python examples/onboarding_china_mirrors.py

The example sets HF_ENDPOINT=https://hf-mirror.com before lazily importing the Hub client and caches the default 44M PII model. To pre-download directly from an installed environment instead:

export HF_HOME="$PWD/openmed-hf-cache"
export HF_ENDPOINT=https://hf-mirror.com
python -c 'from huggingface_hub import snapshot_download; print(snapshot_download("OpenMed/OpenMed-PII-SuperClinical-Small-44M-v1"))'

Archive or copy the entire openmed-hf-cache directory to the offline host, preserving its directory structure and symlinks.

2. Resolve the cache with all Hub traffic disabled¶

On the offline host, point HF_HOME at the transferred directory and enable offline mode before starting Python:

export HF_HOME="$PWD/openmed-hf-cache"
export HF_HUB_OFFLINE=1
export TRANSFORMERS_OFFLINE=1
python examples/onboarding_china_mirrors.py

The example also passes local_files_only=True. If the snapshot is absent, it reports a cache miss instead of attempting a network request. The same flags can wrap an application after its models are cached:

HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 python your_openmed_app.py

See the Hugging Face Hub documentation for the HF_HOME, HF_HUB_CACHE, and HF_HUB_OFFLINE behavior.

PIPL-oriented policy quickstart¶

PIPL compliance depends on purpose, consent or another lawful basis, data classification, retention, cross-border transfer, access controls, and human governance. A software policy profile is only one technical control; this section is not legal advice or a claim of PIPL compliance.

OpenMed ships a china_pipl technical profile for sensitive personal information. It uses the high-recall strict_no_leak threshold profile, requires a safety sweep, replaces direct identifiers, and masks quasi-identifiers and clinical concepts:

from openmed.core.policy import list_policies, load_policy

policy_name = "china_pipl"
assert policy_name in list_policies()

policy = load_policy(policy_name)
assert policy.name == "china_pipl"
assert policy.threshold_profile == "strict_no_leak"
assert policy.safety_sweep_mandatory
assert policy.default_action == "replace"
assert policy.keep_mapping is True
assert policy.reversible_id is True

print(policy.name, policy.default_action)

This snippet runs entirely from the installed package and loads openmed/core/policies/china_pipl.json; it does not need a model or network connection. Pass the same policy_name to the normal runtime call as deidentify(synthetic_text, policy=policy_name). The profile intentionally keeps a reversible mapping for replacement workflows, so its output remains personal information under PIPL; only irreversible anonymization can remove data from PIPL scope. Reassess the policy and your legal controls before production processing of personal information or protected health information (PHI).

Privacy and network boundary¶

Package installers and model-download clients are third-party tools with their own configuration and policies. Keep real personally identifiable information (PII) and PHI out of install commands, model repository names, logs, and mirror requests. Use only synthetic data while verifying this setup.

Troubleshooting¶

• No matching distribution found: confirm the mirror has synchronized the requested OpenMed version, then compare it with the canonical PyPI release.
• LocalEntryNotFoundError in offline mode: the selected model or revision is missing from the transferred cache. Warm that exact snapshot while connected.
• A request still leaves the machine: ensure all environment variables are set before Python starts, and audit any additional application dependencies that have their own network clients.
• A gated or private model fails through the mirror: follow the model owner's access terms and your organization's credential policy; do not place tokens in shell history, source files, or documentation.