HankH18/qwen3-1.7b-forbidden-constraint-holder

Overview

This model, HankH18/qwen3-1.7b-forbidden-constraint-holder, is a QLoRA fine-tune of the Qwen/Qwen3-1.7B base model. Its primary purpose is to improve the model's ability to adhere to "forbidden constraints" during text generation. This includes tasks like writing without using a specific letter or word, where adherence is verified by deterministic checkers rather than subjective judgment.

Key Capabilities and Training

• Constraint Adherence: Specifically trained to hold negative constraints in writing tasks.
• QLoRA Fine-tuning: Utilizes QLoRA (NF4 + double quant) with a LoRA adapter merged to fp16.
• Rejection-Filtered Data: Training data was teacher-distilled and rejection-filtered; completions violating their own constraints were discarded.
• Targeted Improvement: Shows improved compliance on constraint families it was trained on, such as banned_words and no_quotation_marks.
• Prompt Format: Requires a single user turn with no system message for optimal performance, as training was conducted under this condition.

Limitations and Considerations

• Constraint Specificity: The model's performance regresses on constraint families it was not trained on (e.g., banned_punctuation, sentence_initial_letter). If your constraint type is not in the training list, the base model may be a better choice.
• Banned Letter Difficulty: Does not reliably hold banned letter constraints, though it reduces occurrences. A response with any occurrence is still considered a failure by the deterministic checker.
• Prompt Sensitivity: Adding a system prompt puts the model off-distribution, as it was not trained with system prompts.
• Scope: 1.7B parameters, English only, designed for short-form writing tasks.