SmolKalam: Arabic Conversational Supervised Fine-Tuning Dataset

SmolKalam is a quality-filtered Arabic supervised fine-tuning dataset built as an ensemble translation of HuggingFaceTB/smoltalk2. It covers multi-turn dialogue, reasoning traces, tool and function calling, and long-context examples. The dataset was produced by AdaMLLab and last updated on June 22, 2026.

Use Cases

• Fine-tuning Arabic language models for multi-turn conversational agents based on the described dialogue examples.
• Training models to generate reasoning traces (CoT) for Arabic question-answering tasks.
• Developing function-calling capabilities in Arabic AI assistants based on the tool-use examples.
• Benchmarking long-context understanding and generation in Arabic language models.

Strengths

• Focuses on categories underrepresented in Arabic post-training data, including reasoning traces and tool calling.
• Uses a quality-filtering process combining two translation pipelines and selecting the better candidate per sample.
• Last updated on June 22, 2026, suggesting recent maintenance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

AdaMLLab

Collection Method

Ensemble translation of HuggingFaceTB/smoltalk2, combining two translation pipelines and selecting the better of the Gemma 3 and Seed-X candidates per sample.

Freshness

Last updated 2026-06-22 14:57:10.