Turning Language Model Training from Black Box into a Sandbox

Most classroom engagements with generative AI focus on prompting pre-trained models, leaving the role of training data and model mechanics opaque. We developed a browser-based tool that allows students to train a small transformer language model entirely on their own device, making the training process visible. In a CS1 course, 162 students completed pre- and post-test explanations of why language models sometimes produce incorrect or strange output. After a brief hands-on training activity, students’ explanations shifted significantly from anthropomorphic and misconceived accounts toward data- and model-based reasoning. The results suggest that enabling learners to directly observe training can support conceptual understanding of the data-driven nature of language models and model training, even within a short intervention. For K-12 AI literacy and AI education research, the study findings suggest that enabling students to train - and not only prompt - language models can shift how they think about AI.

💡 Research Summary

The paper addresses a gap in AI education: most classroom activities focus on prompting large pre‑trained language models, leaving learners with a black‑box view that obscures how data and model mechanics shape output. To make the training process visible, the authors created the Little Language Machine, a browser‑native tool that runs a small transformer (up to 4 million parameters) entirely on the client using WebGPU/WebGL, mixed‑precision training, and key‑value caching. The implementation is lightweight (≈0.24 GB memory) and GDPR‑compliant, allowing it to run on typical student laptops or Chromebooks without any server interaction.

The study was conducted in a mandatory first‑year CS1 course at a Finnish university. 162 students completed a pre‑test, a hands‑on assignment using the tool, and a post‑test. The pre‑ and post‑tests each contained two open‑ended questions asking why language models sometimes generate incorrect, implausible, or nonsensical text. In the assignment, students selected or uploaded a small text corpus, chose a model configuration, trained the model in the browser, observed loss curves and emergent text, and reflected on how the model’s behavior changed over time.

Student responses were analyzed with qualitative content analysis. An inductive coding scheme produced eight categories: training data issues, model deficiencies, bad prompts, bad web‑search queries, translation errors, lack of intelligence (anthropomorphic view), hallucinations, and others. Before the activity, explanations were dominated by anthropomorphic or erroneous beliefs (e.g., the model searches the web or mistranslates from English to Finnish). After the activity, references to training data and model architecture rose sharply, while misconceptions about web search and translation dropped dramatically.

Quantitatively, mentions of training‑data explanations increased from 13 % (pre‑test) to 38 % (post‑test); a two‑proportion z‑test confirmed this rise as statistically significant (z = 5.09, p < 0.001). At the same time, mentions of “search the web” fell from 37 to 10 instances. Although some students still referenced prompts or hallucinations, these were framed in more technically accurate terms, indicating a shift toward a data‑centric, probabilistic understanding of language models.

The findings demonstrate that even a brief, client‑side training experience can move learners away from common misconceptions and toward disciplinary explanations of model behavior. By turning the opaque training pipeline into an interactive sandbox, the Little Language Machine supports the development of critical AI literacy, aligning with calls for education that treats learners as builders and investigators rather than passive users. The study also highlights the practical advantages of a fully client‑side solution—low cost, privacy preservation, and scalability to classrooms with limited resources.

In conclusion, enabling students to train a language model, not just prompt it, can significantly reshape their mental models of AI. Integrating transparent, hands‑on tools like the Little Language Machine into introductory computing courses offers a promising pathway to more grounded, critical, and inquiry‑driven AI education. Future work should explore longitudinal effects, applicability across K‑12 settings, and how variations in model size, data diversity, and training duration further influence learning outcomes.