Assemble big docket entry dataset

[nadahlberg]

In early experiments the 6.5M entries is not enough data for models to converge. Even with up to 5 epochs. Pull a much larger set of data (from the dev db) for the "real" training runs.