flavianv/qwen3-4b-hm-positive-sft-20260924-v1
The flavianv/qwen3-4b-hm-positive-sft-20260924-v1 model is a 4 billion parameter Qwen3-based causal language model fine-tuned by flavianv. It specializes in recommending coherent clothing bundles for shopping requests, returning JSON arrays of catalog product titles. This model is specifically optimized for generating product recommendations based on H&M catalog data, achieving a micro reference coverage of 5.79% on its development set.
Loading preview...
Overview
This model, flavianv/qwen3-4b-hm-positive-sft-20260924-v1, is a 4 billion parameter Qwen3-based causal language model developed by flavianv. It has been specifically fine-tuned using a public H&M positive-bundle dataset to generate clothing bundle recommendations. The model's training focused on producing JSON arrays of catalog product titles in response to shopping requests.
Key Capabilities
- Specialized Recommendation: Generates coherent clothing bundle recommendations based on H&M catalog product titles.
- JSON Output: Designed to return recommendations strictly as JSON titles arrays, including product name, type, color, pattern, and department.
- Performance: Achieved a micro reference coverage of 5.79% (21/363 products) and successfully matched at least one product for 19 out of 100 queries in its development evaluation.
- Training Details: Fine-tuned over three epochs from an original Qwen3-4B base, updating all backbone weights. Training involved 5,023 positive examples with a completion-only causal token cross-entropy objective.
Usage and Limitations
- Inference Contract: Requires a system message to recommend a coherent clothing bundle and return only JSON with a titles array. Users must specify the required number of distinct products.
- Output Format: Generates at most 512 tokens, with evaluation parameters set to temperature 0.7, top-p 0.9, and top-k 20.
- Known Limitations: Exact product recovery remains low, and the evaluation queries were repeatedly used, meaning they are not an untouched test set. The model returns titles, not authoritative product IDs, and does not include ranking-head training.