Datasets:
run_id stringlengths 22 32 | model stringclasses 4
values | trial int32 1 3 | task_id stringclasses 50
values | task_title stringclasses 50
values | difficulty stringclasses 3
values | instructions stringclasses 50
values | response stringlengths 917 26.4k | all_pass bool 2
classes | criteria_passed int32 0 14 | criteria_total int32 2 15 | criterion_pass_rate float64 0 1 | grades listlengths 2 15 | judge_model stringclasses 1
value | trace_jsonl stringlengths 18.8k 2.61M | trace_event_count int32 8 124 | agent_turns int32 4 56 | tool_calls int32 4 82 | trace_path stringlengths 47 57 | scores_path stringlengths 39 49 | run_metadata_path stringlengths 36 46 | metrics_path stringlengths 43 53 | original_task_path stringlengths 37 47 | grading_task_path stringclasses 50
values | response_sha256 stringlengths 64 64 | trace_sha256 stringlengths 64 64 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
deepseek-v4p1-flash/trial-01/001 | accounts/fireworks/models/deepseek-v4p1-flash | 1 | 001 | FY2025 net revenue reconciliation | easy | "What was FY2025 net revenue, and do SAP and the management accounts agree? Reconcile the reported a(...TRUNCATED) | "# FY2025 net revenue — SAP vs management accounts\n\n**Entity:** Meridian Industrial Supply LLC ((...TRUNCATED) | true | 2 | 2 | 1 | [{"id":"C-001","title":"Reported net revenue","match_criteria":"States FY2025 reported net revenue o(...TRUNCATED) | gpt-6-luna | "{\"sequence\": 1, \"type\": \"model\", \"turn\": 1, \"content\": \"I'll start by exploring the data(...TRUNCATED) | 54 | 22 | 32 | runs/deepseek-v4p1-flash/trial-01/001/transcript.jsonl.gz | runs/deepseek-v4p1-flash/trial-01/001/scores.json | runs/deepseek-v4p1-flash/trial-01/001/run.json | runs/deepseek-v4p1-flash/trial-01/001/metrics.json.gz | runs/deepseek-v4p1-flash/trial-01/001/task.json | grading-tasks/001/task.json | ab9429a0e4c6d5c9cd1297ab8c663bb30b0efeb276a1f9f190d26b375c54f2b2 | b444a7eec2b496ccec1fde3f525634ad6f5256e655e706f8d48054593a4f7cea |
deepseek-v4p1-flash/trial-01/002 | accounts/fireworks/models/deepseek-v4p1-flash | 1 | 002 | Largest customer accounts | easy | "Who were the top five customers in FY2025, and what share of revenue did each have?\n\nOutput: writ(...TRUNCATED) | "# Meridian Industrial Supply LLC — Top five customers, FY2025\n\n## Answer\n\nOn the FY2025 net r(...TRUNCATED) | false | 5 | 6 | 0.833333 | [{"id":"C-001","title":"C412 revenue and share","match_criteria":"Includes C412 (Riverbend Equipment(...TRUNCATED) | gpt-6-luna | "{\"sequence\": 1, \"type\": \"model\", \"turn\": 1, \"content\": \"I'll start by exploring the data(...TRUNCATED) | 55 | 23 | 32 | runs/deepseek-v4p1-flash/trial-01/002/transcript.jsonl.gz | runs/deepseek-v4p1-flash/trial-01/002/scores.json | runs/deepseek-v4p1-flash/trial-01/002/run.json | runs/deepseek-v4p1-flash/trial-01/002/metrics.json.gz | runs/deepseek-v4p1-flash/trial-01/002/task.json | grading-tasks/002/task.json | bd108e5b455d44edcd4e56793f1ed0b57268420eb26a9c192753db8882d95aff | 1628cf25694a063035fbdf3ddb1ec803eafadaa80762204a0d1323d1d96880b8 |
deepseek-v4p1-flash/trial-01/003 | accounts/fireworks/models/deepseek-v4p1-flash | 1 | 003 | Overdue receivables | easy | "How much of year-end AR was over 90 days overdue, and from which customers?\n\nOutput: write your c(...TRUNCATED) | "# Year-end trade AR over 90 days overdue — Meridian Industrial Supply LLC\n\n**Cut-off: 31 Decemb(...TRUNCATED) | true | 2 | 2 | 1 | [{"id":"C-001","title":"Overdue balance","match_criteria":"Identifies C412 (Riverbend Equipment LLC)(...TRUNCATED) | gpt-6-luna | "{\"sequence\": 1, \"type\": \"model\", \"turn\": 1, \"content\": \"I'll start by exploring the data(...TRUNCATED) | 58 | 28 | 30 | runs/deepseek-v4p1-flash/trial-01/003/transcript.jsonl.gz | runs/deepseek-v4p1-flash/trial-01/003/scores.json | runs/deepseek-v4p1-flash/trial-01/003/run.json | runs/deepseek-v4p1-flash/trial-01/003/metrics.json.gz | runs/deepseek-v4p1-flash/trial-01/003/task.json | grading-tasks/003/task.json | 6a299d9cdf653e13794fe20a9fbe261e0e6848990562febfeda85c0b17848ac1 | 4c802fae8212f0a2e35415337af704fbfc36cdd490e29fac52fc9622911d31e6 |
deepseek-v4p1-flash/trial-01/004 | accounts/fireworks/models/deepseek-v4p1-flash | 1 | 004 | Management proposed EBITDA adjustments | easy | "List management's proposed EBITDA adjustments and their total.\n\nOutput: write your complete answe(...TRUNCATED) | "# Management's Proposed EBITDA Adjustments — Meridian Industrial Supply LLC\n\n## Answer\n\nManag(...TRUNCATED) | true | 6 | 6 | 1 | [{"id":"C-001","title":"ERP project","match_criteria":"Lists erp project as a management-proposed ad(...TRUNCATED) | gpt-6-luna | "{\"sequence\": 1, \"type\": \"model\", \"turn\": 1, \"content\": \"I'll start by exploring the data(...TRUNCATED) | 35 | 18 | 17 | runs/deepseek-v4p1-flash/trial-01/004/transcript.jsonl.gz | runs/deepseek-v4p1-flash/trial-01/004/scores.json | runs/deepseek-v4p1-flash/trial-01/004/run.json | runs/deepseek-v4p1-flash/trial-01/004/metrics.json.gz | runs/deepseek-v4p1-flash/trial-01/004/task.json | grading-tasks/004/task.json | 1d6ad2beec288fa47848c3937ceae3d7bd4fcb1525c0d305189d1e8e30b23281 | 2e0a3bee83ab86338a45ba17fc63de10b754671eee9d3577736ccd3bf83f700e |
deepseek-v4p1-flash/trial-01/005 | accounts/fireworks/models/deepseek-v4p1-flash | 1 | 005 | Largest supplier and contract | easy | "Which supplier is largest, what share of purchases does it have, and when does its contract expire?(...TRUNCATED) | "# Largest supplier, share of purchases and contract expiry\n\n**Answer**\n\n- **Largest supplier: A(...TRUNCATED) | true | 2 | 2 | 1 | [{"id":"C-001","title":"Supplier concentration","match_criteria":"Identifies Atlas Motion and Fasten(...TRUNCATED) | gpt-6-luna | "{\"sequence\": 1, \"type\": \"model\", \"turn\": 1, \"content\": \"I'll start by exploring the data(...TRUNCATED) | 53 | 22 | 31 | runs/deepseek-v4p1-flash/trial-01/005/transcript.jsonl.gz | runs/deepseek-v4p1-flash/trial-01/005/scores.json | runs/deepseek-v4p1-flash/trial-01/005/run.json | runs/deepseek-v4p1-flash/trial-01/005/metrics.json.gz | runs/deepseek-v4p1-flash/trial-01/005/task.json | grading-tasks/005/task.json | 743b9520b2b90c1ed940d12215cf2b494f6e182ea7fc382de49ae6d2803a4cfc | 7183105600bbd786e64a9e79ffc443b736c9168a7507c82685ea527647785ba2 |
deepseek-v4p1-flash/trial-01/006 | accounts/fireworks/models/deepseek-v4p1-flash | 1 | 006 | Monthly revenue outlier | easy | "Which FY2025 month is an outlier for revenue, and by how much? Use the January–November monthly m(...TRUNCATED) | "# FY2025 revenue outlier — Meridian Industrial Supply LLC\n\n## Answer\n\n**December 2025 is the (...TRUNCATED) | true | 3 | 3 | 1 | [{"id":"C-001","title":"Outlier month","match_criteria":"Identifies FY2025 December revenue of appro(...TRUNCATED) | gpt-6-luna | "{\"sequence\": 1, \"type\": \"model\", \"turn\": 1, \"content\": \"I'll start by exploring the data(...TRUNCATED) | 43 | 19 | 24 | runs/deepseek-v4p1-flash/trial-01/006/transcript.jsonl.gz | runs/deepseek-v4p1-flash/trial-01/006/scores.json | runs/deepseek-v4p1-flash/trial-01/006/run.json | runs/deepseek-v4p1-flash/trial-01/006/metrics.json.gz | runs/deepseek-v4p1-flash/trial-01/006/task.json | grading-tasks/006/task.json | ed09fed51d9b4fe93e0f556f673b539fa01215763235fed11b5f625e1b277fad | 567559a40988f035721158d96779bf1079abf94fdf0b12362544c62bfa262433 |
deepseek-v4p1-flash/trial-01/007 | accounts/fireworks/models/deepseek-v4p1-flash | 1 | 007 | Closing cash and bank reconciliation | easy | "What cash balance should the net-debt bridge start from at 31 December 2025, and does it agree to t(...TRUNCATED) | "# Cash at 31 December 2025 — starting point for the net-debt bridge\n\n## Answer\n\nThe net-debt (...TRUNCATED) | false | 2 | 3 | 0.666667 | [{"id":"C-001","title":"FY2025 closing cash","match_criteria":"States fy2025 closing cash of USD 8,0(...TRUNCATED) | gpt-6-luna | "{\"sequence\": 1, \"type\": \"model\", \"turn\": 1, \"content\": \"I'll start by exploring the data(...TRUNCATED) | 49 | 19 | 30 | runs/deepseek-v4p1-flash/trial-01/007/transcript.jsonl.gz | runs/deepseek-v4p1-flash/trial-01/007/scores.json | runs/deepseek-v4p1-flash/trial-01/007/run.json | runs/deepseek-v4p1-flash/trial-01/007/metrics.json.gz | runs/deepseek-v4p1-flash/trial-01/007/task.json | grading-tasks/007/task.json | 7d138127eddb8b5df115447f1fe73dc68d5f7592207feda58ccaead67f143de4 | d8d5c3ecb313aaeea7ada166fcc64170219c5eea74535cad65981a98a7e6ea97 |
deepseek-v4p1-flash/trial-01/008 | accounts/fireworks/models/deepseek-v4p1-flash | 1 | 008 | Headcount and payroll | easy | "What were headcount and total payroll cost in each year? Use total recorded payroll expense, includ(...TRUNCATED) | "# Headcount and total recorded payroll cost — Meridian Industrial Supply LLC\n\n## Answer\n\n| | (...TRUNCATED) | true | 2 | 2 | 1 | [{"id":"C-001","title":"FY2024 payroll","match_criteria":"States FY2024 headcount 240 and total reco(...TRUNCATED) | gpt-6-luna | "{\"sequence\": 1, \"type\": \"model\", \"turn\": 1, \"content\": \"I'll start by exploring the data(...TRUNCATED) | 56 | 23 | 33 | runs/deepseek-v4p1-flash/trial-01/008/transcript.jsonl.gz | runs/deepseek-v4p1-flash/trial-01/008/scores.json | runs/deepseek-v4p1-flash/trial-01/008/run.json | runs/deepseek-v4p1-flash/trial-01/008/metrics.json.gz | runs/deepseek-v4p1-flash/trial-01/008/task.json | grading-tasks/008/task.json | 31f1cd57c6dc2653c831154331a7785439c1d0ef2e501c4148ecec5273f2c1e7 | 34e9feb7c7a3e9f3fd28c40f17adc0bf785e1e0b4334ec90f804a7e04d1fbeb2 |
deepseek-v4p1-flash/trial-01/009 | accounts/fireworks/models/deepseek-v4p1-flash | 1 | 009 | Reported EBITDA | easy | "What reported EBITDA do the management accounts show for each year?\n\nOutput: write your complete (...TRUNCATED) | "# Reported EBITDA in the management accounts (Meridian Industrial Supply LLC)\n\n## Answer\n\nThe m(...TRUNCATED) | true | 2 | 2 | 1 | [{"id":"C-001","title":"FY2024 reported EBITDA","match_criteria":"States FY2024 reported EBITDA of U(...TRUNCATED) | gpt-6-luna | "{\"sequence\": 1, \"type\": \"model\", \"turn\": 1, \"content\": \"\", \"tool_calls\": [{\"id\": \"(...TRUNCATED) | 37 | 18 | 19 | runs/deepseek-v4p1-flash/trial-01/009/transcript.jsonl.gz | runs/deepseek-v4p1-flash/trial-01/009/scores.json | runs/deepseek-v4p1-flash/trial-01/009/run.json | runs/deepseek-v4p1-flash/trial-01/009/metrics.json.gz | runs/deepseek-v4p1-flash/trial-01/009/task.json | grading-tasks/009/task.json | 179d1d5078171a38197f4d53998d06e2f1a8de5410af117a8ec2462719484c9e | 037a1d026c689dde245a93ef9fa5624a50a23540ab2c65f0b0ed708089101a27 |
deepseek-v4p1-flash/trial-01/010 | accounts/fireworks/models/deepseek-v4p1-flash | 1 | 010 | Reported gross margins | easy | "What was gross margin in each year?\n\nOutput: write your complete answer to `response.md`, includi(...TRUNCATED) | "# Gross margin — Meridian Industrial Supply LLC\n\n## Answer\n\nGross margin is defined here as g(...TRUNCATED) | true | 3 | 3 | 1 | [{"id":"C-001","title":"FY2024 margin","match_criteria":"States FY2024 reported gross margin of 36.0(...TRUNCATED) | gpt-6-luna | "{\"sequence\": 1, \"type\": \"model\", \"turn\": 1, \"content\": \"I'll start by exploring the data(...TRUNCATED) | 63 | 28 | 35 | runs/deepseek-v4p1-flash/trial-01/010/transcript.jsonl.gz | runs/deepseek-v4p1-flash/trial-01/010/scores.json | runs/deepseek-v4p1-flash/trial-01/010/run.json | runs/deepseek-v4p1-flash/trial-01/010/metrics.json.gz | runs/deepseek-v4p1-flash/trial-01/010/task.json | grading-tasks/010/task.json | 0a7b2533e2c3b6582c8d636b56f50e666ed04e1addeec268f4701d74afda845f | d253fb315a6c4273812fd213e34b0bb86e9690fccf8765772dcd5986042e211e |
End of preview. Expand in Data Studio
FAB — Agent Traces and Grading
600 completed agent runs: four models × 50 tasks × three trials. Agents investigate a synthetic company's data room and answer financial due-diligence questions. Each row pairs a full execution trace with the task, final answer, grading criteria, pass/fail verdicts and judge explanations. Tasks 041 and 049 are included.
The benchmark dataset contains the shared data room and tasks. The GitHub repository contains the execution and grading harness.
Use
import json
from datasets import load_dataset
runs = load_dataset(
"secondstate/finance-agents-benchmark-traces", revision="v1.0", split="test"
)
run = runs[0]
events = [json.loads(line) for line in run["trace_jsonl"].splitlines()]
grades = run["grades"]
trace_jsonl contains the complete original JSONL trace as a string, preserving
provider-specific event fields. grades is a list of criterion IDs, titles,
rubric text (match_criteria), verdicts and judge explanations (reasoning).
The explanations describe grading decisions; they are separate from agent events.
| Fields | Contents |
|---|---|
run_id, model, trial, task_id, difficulty |
Stable run identity and filters |
task_title, instructions, response |
Agent request and original answer |
all_pass, criteria_passed, criteria_total, criterion_pass_rate, grades |
Final grading |
trace_jsonl, trace_event_count, agent_turns, tool_calls |
Full trace and execution counts |
*_path, *_sha256, judge_model |
Original artifacts and provenance |
To download all original artifacts:
hf download secondstate/finance-agents-benchmark-traces \
--repo-type dataset --revision v1.0 --local-dir ./fab-traces
runs/<model>/trial-<NN>/<task>/ contains response.md, final scores.json,
original run.json and task.json, plus transcript.jsonl.gz and
metrics.json.gz. Use gzip -dc <file.gz> to read compressed files.
Final grading rubrics are in grading-tasks/; aggregate results and usage are in
reports. manifest.json records file hashes.
Grading and scope
The judge is gpt-6-luna at maximum reasoning. A task passes only if every
criterion passes. Grades come from the full regrade completed on 27 September
2026, covering all 600 answers; these are existing evaluations, not new runs.
| Model | Task pass rate | Criterion pass rate |
|---|---|---|
| DeepSeek V4.1 Flash | 60.0% | 81.0% |
| GPT-6 Sol | 58.7% | 83.4% |
| GPT-6 Luna | 50.7% | 79.5% |
| GLM 5.3 Flash | 47.3% | 76.2% |
Original run snapshots and final grading rubrics are both retained. Agent
instructions are unchanged; some rubric criteria were clarified before the
full regrade. Grading criteria were kept outside the agent workspace.
Answers, metadata and traces preserve their original bytes; final score files
omit the local source_run path. Failed attempts, scratch files and intermediate
grading checkpoints are excluded. Results describe one synthetic company.
Source: published runs at commit 4c78c7c. License: CC BY 4.0. Cite SecondState's Finance Agents Benchmark and the dataset revision when reusing these results.
- Downloads last month
- 377