Trust and safety

Uncensored LLM for Trust and Safety: A Field Guide

An uncensored LLM for trust and safety is a model with its refusal behavior removed, so it writes and labels the content your systems have to catch. For high volume, use GLM-5.3 Flash. For labels that need careful reasoning, use DeepSeek V4.1 Flash 1017.

You can't teach a classifier to catch scams without showing it scams. Here's how to build that data, test your filters, and what you should never generate.

A field of dark arrows on a fine grid on pale paper, with a few orange arrows running the other way

Why do general-purpose models get in the way?

Ask a filtered model for twenty realistic scam messages and you'll get a lecture about scams. That's the model doing what it was trained to do, and it leaves you with a classifier that has seen nothing new.

Synthetic data can work. The team behind ToxiGen generated 274,000 toxic and benign statements. Human annotators struggled to tell the output from human writing, and fine-tuning a classifier on it substantially improved results on human-written data. With an abliterated model, you can do this work without arguing with the model first.

Which model should you use?

The jobModelEffort
Generating data in bulkGLM-5.3 Flashlow
Labeling against a detailed policyDeepSeek V4.1 Flash 1017high
Finding the gaps in a policyGLM-5.3max

GLM-5.3 Flash has the lowest input price of our models and matches DeepSeek on output, and bulk generation has little to cache. Kimi K3.1 Pro is there for the hardest workloads, and GLM-5.5 is coming soon.

What can you use it for?

  1. GenerateExamples and hard negatives
  2. LabelAgainst your policy
  3. TrainYour classifier
  4. TestOn real, human-written data
  5. RepeatGenerate more where it fails

Training data

Ask for labels and for variety, or you'll get twenty versions of the same message. Placeholders keep real names and links out of your data. Use the same tokens in your real examples, or the classifier just learns to spot placeholders.

Prompt
Here is the scam section of our policy. Write 20 synthetic scam messages, each labeled with the scam type, as training data for our detection classifier. Use placeholder names, links and numbers. Vary the length, tone and pretext. Return JSON.

Hard negatives

Don't skip this step. Classifiers often falsely flag text that mentions minority groups, because those groups are so often the targets of hate. So teach yours the difference with friendly examples.

Prompt
Write 20 ordinary, friendly messages that mention this community, as hard negatives, so our classifier stops flagging people for talking about themselves. Return JSON with the label "benign".

Labeling at volume

Give it your policy, not a vague idea of harm, and let it say when it isn't sure. Those are the ones a person should look at.

Prompt
Here is our policy and 200 reported messages. Label each with the policy section it breaks, or "none", give a one-line reason, and mark any you're unsure of.

Testing your filters

Accuracy on a holdout set hides specific weak spots. HateCheck showed this with 29 functional tests that exposed critical weaknesses in commercial models. You can build the same kind of suite for your own policy.

Prompt
Here is our harassment policy. Write test cases by category, including direct threats, negation, and counter-speech that quotes abuse in order to condemn it. Label each with the decision we'd expect.

Finding the gaps in a policy

Before moderators argue about a case, let the model find the cases worth arguing about.

Prompt
Here is our draft policy. Write 15 borderline examples where two reasonable moderators would disagree, and say which sentence of the policy causes the disagreement.

How do you set it up?

It's a standard OpenAI or Anthropic request. Change the base URL and use a Silk Compute key.

Python
import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.silkcompute.com",
    api_key=os.environ["SILK_API_KEY"],
)

response = client.chat.completions.create(
    model="glm-5.3-flash-uncensored",
    reasoning_effort="low",
    messages=[{"role": "user", "content": """Here is the scam section of our policy. Write 20 synthetic scam messages, each labeled with the scam type, as training data for our detection classifier. Use placeholder names, links and numbers. Vary the length, tone and pretext. Return JSON."""}],
)

print(response.choices[0].message.content)

To work in a coding agent instead, set it up for Claude Code, Codex or OpenCode. Every model works with tool calling, so the agent can read files and run commands as usual.

What should you never generate?

Never generate sexual content involving anyone under 18, for any reason, including training data. It's the one rule in our Terms: no sexual content involving anyone under 18, and no using the models to abuse or exploit a child.

Detecting that material doesn't call for generated examples. Programs such as Safer by Thorn offer hash matching for known material and specialist classifiers. In the US, suspected material gets reported to the NCMEC CyberTipline. Talk to your counsel about your obligations.

How do you keep the data honest?

Synthetic data can help scam and toxicity classifiers, but it's easy to fool yourself with it:

  • Test on real data. Keep a holdout set of human-written examples, and judge the classifier on that. One 2024 study across six toxicity datasets found synthetic data helped in some scenarios and did not improve hate speech classification.
  • Filter what you generate. ToxiLab scored six open models on hallucination and duplication, because both show up in this kind of data. Dedupe and review before you train.
  • Tag everything synthetic. You'll want to know later which examples a person actually wrote.
  • Check the labels. Sample them by hand, and measure how often a person agrees.
  • Look after your people. Generated abuse is still abuse to read. Rotate reviewers and limit how much anyone sees in a day.

Prompts and model outputs pass through our servers and are not stored. We keep only billing records, such as token counts and cost.