BlazingCustoms/acrostic-4b-v4a
BlazingCustoms/acrostic-4b-v4a is a 4 billion parameter LoRA fine-tune of Qwen/Qwen3-4B-Instruct-2507, specifically designed to generate acrostic poems in a position-anchored format. This model excels at creating acrostics for target words of lengths 3-10, achieving 96.3% accuracy across 380 items with a worst-case accuracy of 85.0% at length 8. It is optimized for precise acrostic generation, requiring an external oracle for verification to ensure correctness.
Loading preview...
Model Overview
BlazingCustoms/acrostic-4b-v4a is a 4 billion parameter LoRA fine-tune of the Qwen3-4B-Instruct-2507 base model, specialized in generating acrostic poems. Unlike its 3B predecessor, this model leverages the stronger Qwen3-4B base, significantly improving performance on longer acrostics, particularly for out-of-distribution lengths up to 10.
Key Capabilities
- Acrostic Poem Generation: Creates acrostic poems where the first letter of each line spells a target word, presented in a numbered, labeled format (e.g., "1. S | Surging waves...").
- High Accuracy: Achieves 96.3% overall accuracy on a 380-item held-out test set for target lengths 3-10 at temperature 0. The model's worst performance is 85.0% at length 8, with length 10 (out-of-distribution) reaching 98.0% accuracy.
- Robustness: The fine-tune improves general instruction-following by 3.0 percentage points and retains 99.3% lexical diversity compared to the untuned base, indicating no mode collapse.
- Quantized Builds: Offers GGUF quantized builds (Q6_K, Q5_K_M, Q4_K_M) that are statistically indistinguishable from the f16 model in accuracy, providing options for reduced memory footprint.
Important Considerations
- Verification Oracle Required: The model's safety and reliability for acrostic generation depend on an external, deterministic verification oracle (provided in the README) to check and resample outputs. Without this, silent failures (incorrectly spelled words) can occur.
- Limitations: Untested above length 10 or below length 3. Poem quality is unmeasured and not claimed. Rare initial letters (X, K, Q, Z) are the primary lexical failure mode.
Good For
- Applications requiring precise acrostic poem generation with a strong emphasis on the first-letter constraint.
- Developers looking for a specialized, high-accuracy model for a niche creative writing task.
- Use cases where quantized models are preferred for local or server-side inference, with GGUF builds available.