Two mislabeled gold answers in dev set (AOS/2004, VLO/2018)

[ashalliants]

While working with ConvFinQA I found two conversation turns in the dev set whose gold execution answers don't match the question being asked. In both cases the discrepancy is traceable to the turn-collapse/decomposition step, and the annotation fields in the same record contain the evidence for the correct answer.

1. Double_AOS/2004/page_11.pdf, turn 0

• Question: "as of december 31, 2004, what percentage of the total of aggregate contractual obligations was due to long-term debt?"
• Current turn_program[0] = "275.1", exe_ans_list[0] = 275.1 — this is the bare numerator, not a percentage.
• The same record's original_program_0 = divide(275.1, 521.3) and step_list_0 = ["Ask for number 275.1", "Ask for number 521.3", "divide(275.1, 521.3)"]. When the two "n/a" ask-for-number turns were collapsed (dialogue_break_ori → dialogue_break), turn 0's program was aligned to the intermediate 275.1 step instead of the final divide.
• Correct answer: divide(275.1, 521.3) ≈ 0.5277 (52.77%). This is also consistent with the parallel turn 1 ("and what percentage of it was due to purchase obligations?" → divide(177.3, 521.3)).

2. Double_VLO/2018/page_25.pdf, turn 0

• Question: "what was the total number of shares purchased in november for the q4 ended 12/31/18?"
• Current turn_program[0] = divide(3655945, 7673266), exe_ans_list[0] = 0.47645 — a ratio, but the question asks for a count.
• The table value for November total shares purchased is 3,655,945, which is the answer to the question as phrased. (The divide(...) ratio would answer a "what percentage of total" question, not this one.)

Both reproduce in train.json / dev.json from the current data.zip. Happy to provide a small patch/errata file if useful.