psytechlab/Qwen3-4B-Instruct-2507-lora-depressive_style

DeproLLM — Qwen3-4B GRPO-tuned for Depressive-Style Essay Generation

Model Description

DeproLLM is a fine-tuned adapter for Qwen3-4B-Instruct-2507, trained to generate essays that reproduce the stylistic markers of depression rather than merely its topical content. The model was trained with GRPO (Group Relative Policy Optimization) on top of a LoRA (Low-Rank Adaptation) adapter.

The core motivation is methodological. Work at the intersection of machine learning and psychology is hard because ML thrives on large labeled datasets, while clinical data on individuals' mental-health problems is scarce, sensitive, and typically restricted to a single study. A prior dataset used for classical ML depression classification contained only 316 essays, of which just 93 belonged to the target (diagnosed-depression) class — enough for feature-based analysis, but too small and too narrow (one genre, one prompt) to train and evaluate flexible models across diverse texts. In fact, this dataset is closed, the only the clasifier is available. This is a motivation to use the GRPO.

The detalied description is available in this post.

Intended Use

• Primary use: Generating synthetic essays with depressive stylistic characteristics to augment scarce, privacy-restricted training data for downstream depression-detection research.
• Not intended for: Clinical diagnosis, medical decision-making, or any real-world screening of individuals. This is a research artifact.

Training Approach

Reward model / judge

Rewards during GRPO were derived from a depression-text classifier from the TITANIS toolkit:

1. Extract 73 psycholinguistic features
2. Scale with StandardScaler
3. Reduce dimensionality via PCA
4. Classify with an SVC (support vector classifier), which outputs a continuous score.

This continuous score makes the classifier suitable as a judge inside the GRPO loop.

Reward shaping (anti-hacking)

Optimizing purely against the classifier risks degenerate solutions — e.g., collapsing essays to minimal length to inflate marker density, or repeating markers. To prevent reward hacking, additional reward components were added based on:

• Text length
• Repetition penalties
• A penalty for explicit depressive vocabulary (discouraging shortcuts through overt lexical content rather than genuine stylistic modeling)

Evaluation

The generation target metric was the TITANIS classifier's score. DeproLLM was compared against:

• the base model with various prompting strategies, and
• prompting of a larger available model.

Using the classifier score as the metric is a known limitation.

Limitations

• Genre scope: Training focused on the essay format, matching the domain the classifier was trained on. Other domains (e.g., social-media posts) are left for future work.
• Metric dependence: Evaluation relies on a single classifier, so the model may reflect that classifier's biases.
• Synthetic-only: Outputs are artificial and should not be treated as representative of any real individual's mental state.

Ethical Considerations

This model concerns a sensitive mental-health domain. It must not be used to diagnose, profile, or make consequential decisions about real people. Any research use should follow appropriate ethical review, and outputs should be handled with care to avoid harmful or misleading applications.

LLM usage disclosure

The model card was generated by LLM based on the blog post that describes the results. The generated text was reviwed by a human.