This paper details the pjmathematician team's participation in the TalentCLEF 2025 shared task, focusing on Task A (Multilingual Job Title Matching) and Task B (Job Title-Based Skill Prediction). Our approach primarily leveraged state-of-the-art sentence embedding models fine-tuned using the GISTEmbed technique. For Task A, various multilingual and English-specific encoder models were adapted, including a distilled version and a LoRA-fine-tuned 7B parameter model. Data augmentation for Chinese was performed using Qwen2.5 32B Instruct. For Task B, we employed data augmentation using Qwen2.5 32B Instruct to generate descriptive texts for jobs and skills, significantly enriching the training data. Models like BGE-Large and GTE-Qwen2-7B (LoRA) were ifne-tuned on this augmented data. Our submissions demonstrate the efectiveness of these strategies, achieving competitive results in both tasks.
Our approach for Task A centered on fine-tuning various sentence embedding models to capture semantic similarities between job titles across multiple languages.
We experimented with several base encoder models:
• BAAI/bge-small-en-v1.5 (33.4M parameters) [
All models, except the distilled one, were fine-tuned using the GISTEmbed loss [
All fine-tuning and inference were performed using the ‘sentence-transformers‘ library [
For Task B, our strategy focused on robust data augmentation using a Large Language Model (LLM) and fine-tuning powerful English-specific sentence encoders.
We utilized the Qwen2.5 32B Instruct (AWQ quantized) model to generate descriptive text for both job titles and skills. The process involved prompting the LLM to create concise descriptions: • For jobs: "Given a job role (and its synonyms), briefly (1-2 sentences) describe the skills needed for that job". This generated a ‘skill_brief‘ for each job title. • For skills: "Given a skill (and its synonyms), briefly (1-2 sentences) describe the job roles that require that skill". This generated a ‘job_brief‘ for each skill.
This process was applied to the training, validation, and test sets provided by the organizers. Two main augmented training datasets were created from the original job-skill mapping files: 1. A dataset of pairs (‘skill_brief‘, ‘job_brief‘). 2. A dataset of mixed-format pairs (‘skill_brief‘ + newline + job synonyms list, ‘job_brief‘ + newline + skill synonyms list).
We employed two main base models:
• BAAI/bge-large-en-v1.5 (335M parameters) [
The ‘sentence-transformers‘ library [
The models were evaluated based on Mean Average Precision (MAP) as the oficial metric.
We submitted five systems for Task A, varying the base model and fine-tuning techniques. The GTEQwen2-7B model fine-tuned with LoRA and CachedGIST achieved the best overall performance across English, Spanish, and German with an Avg.MAP of 0.52. The results are summarized in Tables 1, 2, and 3.
All models were fine-tuned using the sentence-transformers library with a batch size of 64 and the AdamW optimizer. We used the CachedGISTEmbedLoss to align model outputs with embeddings generated from the all-MiniLM-L12-v2 guide model. The training ran for 1–3 epochs with early stopping based on MAP scores on a 10% validation split.
For the GTE-Qwen2-7B model, we applied LoRA fine-tuning with a rank of 8 and trained only adapter layers to reduce GPU memory requirements. For the distilled version of gte-multilingual-base, we used a teacher–student setup, training the student with MSE loss to mimic embeddings from the GISTEmbed-fine-tuned full model.
At inference time, both query and candidate job titles were encoded using the fine-tuned model, and cosine similarity was computed to rank candidate titles. We used the same multilingual model per base encoder across all supported languages, without applying any task-specific heuristics. For Chinese job titles, LLM-generated translations of ESCO data ensured consistent structure and terminology with the English training examples.
A major challenge in Task A was the lack of labeled training data for Chinese. To overcome this, we translated English job titles from ESCO into Chinese using the Qwen2.5 32B Instruct model. The motivation was to create aligned examples that preserved semantic structure while leveraging a powerful LLM’s cross-lingual generation capabilities. By training a single multilingual model for all languages (en, es, de, zh), we aimed to ensure consistent semantic space alignment and reduce the complexity of maintaining separate models.
This design choice enabled the model to learn language-agnostic representations of job titles, facilitating strong cross-lingual performance as reflected in the MAP scores.
For Task B, we submitted three systems, leveraging LLM-augmented data. The results are shown in Table 4.
The GTE-Qwen2-7B model fine-tuned with LoRA on the mixed augmented data (descriptions and synonym lists) yielded the highest MAP of 0.36. This suggests that the combination of a large instructiontuned base model, LoRA, and rich augmented input was most efective. The BGE-Large model trained solely on the augmented descriptions performed competitively (MAP 0.34). When BGE-Large was trained on the mixed augmented data, the performance was slightly lower (MAP 0.33). This might indicate that for BGE-Large, the simpler augmented descriptions were more directly beneficial, or that the mixed data format required further hyperparameter optimization. The use of LLM-generated descriptions proved crucial for providing rich textual context.
All models were fine-tuned using the sentence-transformers library with a batch size of 64, using the AdamW optimizer and a cosine learning rate schedule. We employed early stopping based on validation MAP. For GISTEmbed-based training, the guide model was set to all-MiniLM-L12-v2, and the CachedGISTEmbedLoss was used to align the student model’s embeddings with cached guide embeddings.
For the GTE-Qwen2-7B model, we used Low-Rank Adaptation (LoRA) with a rank of 8 and bias training enabled, to eficiently fine-tune the model without updating all parameters. Fine-tuning was performed for 1–3 epochs depending on convergence behavior, monitored using a 10% validation split from the training set.
During inference, we used the same augmentation templates as during training. Each job title query was transformed into a prompt-generated description (optionally combined with a list of aliases), and similarly for each skill. These texts were encoded into embeddings using the fine-tuned model, and cosine similarity was computed between the job title and all skills. The top-N most similar skills were returned as the model’s output for ranking.
In mixed-format submissions, we used newline-separated concatenations of generated descriptions and alias lists. This format provided the model with richer and more consistent context and led to improved generalization.
The core motivation for data augmentation was to enrich the semantic content of both job titles and skills, which are otherwise short and ambiguous. By prompting the Qwen2.5 32B Instruct model to generate compact but expressive descriptions, we aimed to reduce lexical sparsity and improve the model’s ability to match based on conceptual similarity.
Additionally, including known aliases in the input helped align representations across synonymous phrases. The combination of these strategies allowed the model to learn more generalizable embeddings, making it better suited for real-world applications where job titles and skill names vary significantly in wording.
In our participation in TalentCLEF 2025, we explored the eficacy of fine-tuning sentence embedding models using GISTEmbed for both multilingual job title matching (Task A) and job title-based skill prediction (Task B). For Task A, larger models like GTE-Qwen2-7B, fine-tuned with LoRA, demonstrated superior performance, particularly when augmented with LLM-translated data for low-resource languages like Chinese. For Task B, data augmentation via LLM-generated job and skill descriptions was a key strategy. The GTE-Qwen2-7B (LoRA) model trained on mixed augmented data (descriptions and synonyms) achieved the best results, underscoring the value of rich, contextualized training inputs. Our experiments highlight the potential of combining advanced fine-tuning techniques like GISTEmbed and LoRA with LLM-driven data augmentation for complex semantic matching tasks in the HR domain.
During the preparation of this work, the author(s) used Qwen2.5 32B Instruct in order to: Generate training data through textual descriptions for jobs and skills (Task B), and Translate English job titles to Chinese for training data augmentation (Task A). After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.