metaphilabs/credit-underwriting-preview

Credit Underwriting Benchmark

A 1387-task RL environment for financial document understanding. Agents are provided with business documents (bank statements), and asked 
to assess the risk associated by using a context-rich classification taxonomy for each transaction . The agent's output is programmatically graded 
against the ground truth labelled by domain epxerts with over 10+ years of experience in the risk assessment field. 

Scale

• 1387 qualifying cases (Businesses, loan-origination)
• 5,607,517 total transactions
• 1-73 PDFs per case (median: 4, mean: 6.4)

Task Definition

Input	1-73 bank statement PDFs per case, no PII information
Output	Extract Structured JSON: transactions (date, amount, balance, narration) and categorize transaction type, counterparty
Reward	Extraction F1 (hash-based), categorization accuracy
Grading	Fully programmatic, instant (~seconds per case)

Difficulty Distribution

By PDF Count

PDFs	Cases	% of Total
1	269	19.4%
2-3	406	29.3%
4-6	317	22.9%
7-10	154	11.1%
10+	241	17.4%

By Transaction count

Transactions	Cases	% of Total
<100	11	0.8%
100-500	153	11.0%
500-2,000	535	38.6%
2,000+	688	49.6%

Financial Data Analyst (Extraction)

Extracts structured transaction data from PDF bank statements.

Model	Harness	Pass@1	95% CI	Tasks
GPT-5.4	Codex	33.3%	8.3 - 58.3%	12
Claude Opus 4.6	Claude Code	31.3%	12.5 - 56.3%	16
Gemini 3.1 Pro	Gemini CLI	12.5%	0.0 - 29.2%	12

Credit Analyst (Categorization)

Categorizes bank transactions by type, counterparty, and group.

Model	Harness	Pass@1	95% CI	Tasks
Claude Opus 4.6	Claude Code	31.3%	12.5 - 56.3%	16
GPT-5.4	Codex	16.7%	0.0 - 41.7%	12
Gemini 3.1 Pro	Gemini CLI	12.5%	0.0 - 33.3%	12

For RL Training

• Reward signal: Extraction F1 as scalar reward (0.0 to 1.0), fully programmatic
• Rollout: Agent receives PDFs → produces structured JSON → verifier computes F1 in seconds
• Parallelizable: Each case is independent — runs in an isolated container
• Difficulty calibration: Easy (84% F1) to hard (43% F1) — significant headroom
• No human evaluation needed: reward signal is instant and deterministic
• Decomposed rewards available: extraction F1 + categorization accuracy + counterparty accuracy as separate signals

Dataset Schema

```json
{
  "id": "case-000",
  "case_id": "uuid-based-case-identifier",
  "pdf_gcs_uris": ["gs://.../_scrubbed.pdf"],
  "pdf_filenames": ["fileid_scrubbed.pdf"],
  "extraction_ground_truth": "{...}",
  "categorization_ground_truth": "{...}",
  "description": "3 PDFs - 792 txns",
  "case_type": "Business Case",
  "gt_status": "clean"
}
```