Where the gains clustered
Particle-1’s gain in data analysis becomes clearer when we look at where the questions were difficult. In Plotly’s 43-question exam it passed 40, against 36 for its DeepSeek V4.1 Flash base. The models tied on ordinary analytics, ambiguous requests and impossible requests. The difference came from statistical traps and patterns hidden in the data. That distribution is a more useful starting point than the total: answering an everyday query and questioning a misleading aggregate should be treated as different skills.[1]
Plotly used the same software setup and grading rule for the models. Its warehouse represents a synthetic wind-energy operator, designed so answers cannot simply be recalled from familiar real datasets encountered in training. That choice narrows the route of remembering a known dataset’s conclusion. It also leaves much of the difficulty of an organization’s messy data outside the test. The evaluation shows behavior within a bounded analytical task; extending the result to the full range of business analysis requires additional comparisons.[1]
When the training objective meets another task
The interesting result is that the training objective did not transfer intact. Particle-1’s extra training targeted fewer tokens, the pieces used to process text, in software work. On Plotly’s exam it instead produced about 6,900 output tokens per question, against about 6,200 for the base. Model calls rose from roughly 7 to 10. More correct answers did not arrive with shorter answers here. Assuming that efficiency learned on one task must carry into another would assume the very outcome that needs measuring.[1]
This does not establish that the adaptation failed. Extra computation may have helped the model check difficult questions. An alternative is that it took longer paths through the same tasks and some of the observed difference came from variation between runs. Plotly says repeats with other models have differed by as many as three questions. A single run cannot separate those contributions. Repeated, paired comparisons on the same questions should track both correct answers and token use.[1]
Counting false alarms too
The error type should remain visible beneath the total. During a data-quality audit Particle-1 treated a pattern explained by other data as a contradiction and advised against trusting those rows. The base model answered that question correctly. Additional correct answers can therefore coexist with a new false alarm. For a researcher or analytical team, a useful evaluation exposes unnecessarily discarded data alongside detected patterns. Two systems with the same total can leave different work for the person reviewing their output.[1]
The next comparison should distinguish tasks before expanding the league table. Reporting accuracy alongside computational use for ordinary queries, statistical traps and data-quality problems would show more clearly what the adaptation is useful for. Publishing the distribution across repeat runs would also show how stable a difference of a few questions is. That design would help model selection without automatically equating brevity with efficiency or extending an aggregate score to every analytical job.[1]