Training Data

The quality and diversity of training data directly affects the reliability of model outputs, and training data containing large amounts of low-quality, contradictory, or false information can cause models to learn incorrect patterns.

Open-source licenses guarantee the availability of code and weights, but the composition of training data, details of the training process and the model's internal decision logic remain a black box to external users.

In the first round of training data, 86% of rows were labeled allow, and the median length of real allow commands was 287 characters, whereas hand-written risky commands were usually only 33 to 40 characters, causing the model to learn the spurious signal that "short command = dangerous".

The training data consists of about 13,000 shell commands from the author's own Agent conversation records, of which 696 commands not covered by deterministic rules were manually labeled; commands with a similarity of no less than 0.85 were excluded.

The third round of fine-tuning reduced the allow proportion in the training set to 61%, achieved by removing 242 safe samples already covered by deterministic rules, duplicating review/block examples, and adding 100 targeted examples.

After the second round of fine-tuning, three dangerous command cases were missed: rmdir /s /q d:, a destructive command after a comment, and gh auth token; the reasons were respectively that the training data lacked Windows commands, that comment samples skewed toward allow, and that there were no samples for printing keys.

If a model learns with police stop records over the years as labels, its output is essentially a reproduction of existing law enforcement behavior rather than an objective prediction of unlawful behavior.

This genre is a recursive mixture of prior conversational discourse between humans and generative models, and new text produced by human-machine dialogue may be incorporated into subsequent training data.

The research team used a generative video editor to place animated 3D assets into matching scenes and filter the edited results, combining human and robot videos to ultimately produce nearly 200,000 training pairs.

Different models vary in their sensitivity to cultural localization, with the magnitude of decline differing by model and closely related to the language distribution and cultural coverage of the training data.

After the researchers reformulated explicit criteria, 38 problematic questions were ultimately excluded from supervision, and the fixed training used 12,520 records and 15,476 questions.

The output of automatic tracking is being used as training data, so tracking errors are no longer an experience issue but a data contamination issue.

For detecting identity switches before training, there is currently no standard answer or cheap verification method.

Painting with code requires a large number of 'code-image' paired training samples, which are scarcer than 'text-image' pairs.

High-quality training data available on the public internet is being rapidly consumed by large models and is becoming scarce.

Models unconsciously inherit the values, biases, and expression habits present in their training data.

The quantity of high-quality publicly available underwater whale videos is far lower than the volume of data available for common terrestrial scenes.

AI naturally tends to generate the most common styles in its training data; if the training corpus is dominated by Western classical aesthetics, the renaissance may devolve into neoclassical commodity output.

Legally licensed, high-quality training data may in the future become a core competitive barrier among AI companies