Finance Agent Sample Set
Build financial work that survives changed inputs.
14 tasks · 6 finance workflows · 2 reference solutions + agent traces · v2.7.1
Source artifacts, linked workbooks, and structured outputs for financial-agent evaluation. Test end-to-end consistency, input responsiveness, and numerical causality under task-specific scoring.
Inspect the work. Trace the failure. Measure the result.
Sample Gallery · Contact · Evaluation terms
Evidence before credit
final_score = min(natural_score, applicable_score_ceiling)
Public files support artifact inspection. The complete verifier, replay runtime, and scoring receipts belong to the reviewer package.
Retained results
14 tasks × 3 retained slots × 2 models. Scores / 100.
Cohort caveat: Sol was selected for valid submissions with natural score < 50. Astra had no score threshold and allowed invalid-submission replacement before freeze. These are historical sample outcomes, not an unbiased model ranking or evidence of training lift.
Task-level score distribution
Each observation is a task mean over three slots; headline means weight tasks equally. Full credit = 100. Click either results figure for full resolution.
Score by workflow
84 slot scores · Aggregation and selection
Inspect typical failure cases
Astra's retained-slot median is 25/100; 18 of 42 slots score exactly 25. These two cases show common failure patterns: numerical causality and financial controls.
| Case | Final score | Failure pattern | Inspect |
|---|---|---|---|
| DCF Mall · slot 01 | 25.00 | Numerical causality; 25-point cap | Reference · Model · Trace |
| M&A 0040 · slot 02 | 38.64 | Financial controls; natural score below 49-point cap | Reference · Model · Trace |
DCF Mall sits at the cohort median; slot 01 breaks its score tie. M&A 0040 slot 02 is the financial-control-failure case closest to the cohort mean (37.80). This is purposive case selection, not a statistically representative sample. Selection and provenance · All 42 Astra slots.
Reference workbooks are separate from model submissions. JSON answers accompany both. Traces retain agent actions and observations; bootstrap prompts and run metadata are omitted. Historical scores are unchanged.
Browse all 14 tasks
DCF & operations · LBO & SEC retrieval · Cash flow · Loan tape · M&A & funding · IRR & comparables
Open task instructions and submission notes
| Task | Start here | Supporting documentation |
|---|---|---|
| DCF_LNG_020 | Instructions | Data and submission notes |
| DCF_MALL_030 | Instructions | Data and submission notes |
| LBO_DAR_001 | Instructions | Data and submission notes |
| LBO_DAR_001_STAGE2 | Instructions | Data and submission notes |
| LBO_ACM_021 | Instructions | Data and submission notes |
| LBO_ACM_021_STAGE2 | Instructions | Data and submission notes |
| CFS_S2E2_012 | Instructions | Data and submission notes |
| CFS_S2E2_020 | Instructions | Data and submission notes |
| LOAN_ELIGIBILITY_01 | Instructions | Data and submission notes |
| LOAN_BORROWING_BASE_05 | Instructions | Data and submission notes |
| MA_COMBO_0040 | Instructions | Data and submission notes |
| MA_COMBO_0824 | Instructions | Data and submission notes |
| IRR_B08 | Instructions | Data and submission notes |
| IRR_R05 | Instructions | Data and submission notes |
Start with instruction.md, then the task's data and output contracts. /app paths refer to the container environment; SEC retrieval tasks need network access. Public validators check submission interfaces, not complete scoring.
Task files are byte-identical to the v2.7.1 candidate package. File manifest and hashes.
Access and contact
Broader task sets, training access, or a scoped evaluation pilot: J4redleung@gmail.com.
Technical evaluation is permitted under Evaluation Sample Terms. Training, fine-tuning, redistribution, and commercial use beyond internal evaluation require written permission.
- Downloads last month
- 264