zou-group / humanlm

What is humanlm?

HumanLM is a Python toolkit for building datasets to train language models that simulate specific users by aligning to their internal states—like beliefs and stances—rather than just imitating surface responses. It scrapes real conversations from sources like Reddit, YouTube comments, Amazon reviews, Medium politics blogs, AI chats, and Enron emails, then processes them into Hugging Face-ready splits with LLM-generated user personas. Developers get clean train/val/test sets emphasizing humanism and state alignment, which beats plain response imitation in user studies.

Why is it gaining traction?

It delivers diverse, multi-turn datasets with automatic persona summaries from a user's comment history, enabling more consistent human-like simulation across domains. The scrapers handle APIs robustly with retries and proxies, while processing flags outliers and partitions data to prevent leakage—features that save weeks of manual work. Early evals show state-aligned models outperforming imitation baselines, hooking researchers chasing realistic user behaviors.

Who should use this?

AI researchers fine-tuning LMs for personalized agents or alignment benchmarks, especially those simulating forum users, email threads, or chat responses. It's ideal for teams evaluating how well models capture individual opinions without overfitting to generic replies, like in debate bots or virtual personas.

Verdict

Grab it if you need battle-tested data pipelines for user simulation—docs and scrapers are solid despite 44 stars and 1.0% credibility score. Skip for production until training code lands; it's raw tooling, not a full framework.

(178 words)