An experiment tested how temperature settings and random seeds affect the output of open-source AI models. The researcher used a prompt about a museum night guard finding a warning note, requesting a five-sentence story with a specific reveal.
The test involved two models, Qwen2.5-1.5B-Instruct and Qwen3-0.6B, across three temperature levels: 0.2, 0.7, and 1.2. For each setting, eight different random seeds were applied to generate 48 distinct responses.
At a temperature of 0.2, Qwen2.5 produced varied narratives, including different authors for the note, while Qwen3 remained more consistent in its storytelling. Higher temperatures led to increased narrative divergence and occasional loss of coherence.
At a temperature of 1.2, both models exhibited significant degradation. Responses included fragmented text, mixing of different languages, and the insertion of non-textual data, with some outputs failing to complete within the 240-token limit.
