Prompt Clipboard — thisisphume

FILES

HI-Small FS — Batch Run Data

system prompt + 500 user prompts + ground truth manifest

File

Size

Description

system_prompt.txt

122 KB

Preamble + 12 few-shot examples (8 fraud + 4 legit)

Download

16 user_prompt files (250 cases each, click to expand)

user_prompts_00.jsonl	19.8 MB	Cases 1–250	DL
user_prompts_01.jsonl	17.4 MB	Cases 251–500	DL
user_prompts_02.jsonl	17.3 MB	Cases 501–750	DL
user_prompts_03.jsonl	17.8 MB	Cases 751–1000	DL
user_prompts_04.jsonl	16.4 MB	Cases 1001–1250	DL
user_prompts_05.jsonl	18.8 MB	Cases 1251–1500	DL
user_prompts_06.jsonl	22.5 MB	Cases 1501–1750	DL
user_prompts_07.jsonl	17.0 MB	Cases 1751–2000	DL
user_prompts_08.jsonl	17.2 MB	Cases 2001–2250	DL
user_prompts_09.jsonl	16.5 MB	Cases 2251–2500	DL
user_prompts_10.jsonl	16.7 MB	Cases 2501–2750	DL
user_prompts_11.jsonl	14.4 MB	Cases 2751–3000	DL
user_prompts_12.jsonl	19.5 MB	Cases 3001–3250	DL
user_prompts_13.jsonl	17.3 MB	Cases 3251–3500	DL
user_prompts_14.jsonl	15.3 MB	Cases 3501–3750	DL
user_prompts_15.jsonl	0.6 MB	Cases 3751–3753	DL

manifest.json

425 KB

Ground truth: case_id → label + typology (all 3,753 cases)

Download

system_prompt_ZS.txt

0.2 KB

Zero-shot preamble (no few-shot examples)

Download

task_suffix_ZS.txt

0.5 KB

ZS task instructions + JSON output schema (appended after graph)

Download

run_gemini_enterprise_py.txt

7 KB

Vertex AI runner — loads JSONL, calls Gemini, saves results

Download

compute_metrics_py.txt

7 KB

Parse responses → Det-F1, Typ-F1, etc.

Download

3,753 cases in 16 files (under 23 MB each, Cloudflare Pages safe). Full V2 pool = ~$680 on Gemini 2.5 Pro (FS). Start with just _00 (250 cases, ~$45) and scale up.

suspicious

v2_02992 — fan-in, 5 txns

ground truth: illicit | 2,369 chars

You are an expert financial crime investigator reviewing patterns of financial activities and behaviors of involved accounts to identify potential cases of money laundering. The data is represented as a graph, where:
- Nodes are of type Account or Bank.
- Edges represent relationships of type transfers_to or belongs_to, and include metadata such as amount, currency, payment method, and timestamp.

For training purposes, you will be shown examples of subgraph typologies that are known to be either suspicious (indicative of laundering tactics) or non-suspicious (routine financial activity). These typologies illustrate common structural patterns in financial networks.

Task: Given a transaction (edge) with Transaction ID <ID>, along with its surrounding subgraph, determine whether the transaction is suspicious, or not suspicious. Your reasoning should be based on whether the surrounding subgraph resembles any of the suspicious typologies provided in the examples.
Test Example:
=== Transaction Subgraph (Case: v2_02992) ===

**Nodes:**
- acct_80EF87ED0 (type: Account)
- acct_80EF86180 (type: Account)
- acct_80EF861D0 (type: Account)
- acct_80EF87C60 (type: Account)
- bank_0140257 (type: Bank)
- bank_018067 (type: Bank)
- bank_019477 (type: Bank)
- bank_039322 (type: Bank)

**Edges:**
- acct_80EF87ED0 belongs_to bank_0140257
- acct_80EF86180 belongs_to bank_018067
- acct_80EF861D0 belongs_to bank_019477
- acct_80EF87C60 belongs_to bank_039322
- acct_80EF861D0 transfers_to acct_80EF861D0 amount: 14356.01 Yuan via: Reinvestment timestamp: 2022/09/01 00:16
- acct_80EF87C60 transfers_to acct_80EF861D0 amount: 3092.84 Yuan via: Cheque timestamp: 2022/09/02 10:35
- acct_80EF87C60 transfers_to acct_80EF861D0 amount: 3092.84 Yuan via: Cheque timestamp: 2022/09/09 01:50
- acct_80EF87ED0 transfers_to acct_80EF861D0 amount: 313378.36 Yuan via: ACH timestamp: 2022/09/10 04:53
- acct_80EF86180 transfers_to acct_80EF861D0 amount: 71391.54 Yuan via: ACH timestamp: 2022/09/10 04:59

SUMMARY STATISTICS:
  Total transactions: 5
  Unique accounts: 4
  Total volume: 405,311.59
  Avg transaction: 81,062.32
  Time span: 5 units
  Max out-degree: 2 (node: 039322_80EF87C60)
  Max in-degree: 5 (node: 019477_80EF861D0)
Answer Format:
- Conclusion: Suspicious or Not Suspicious
- Explanation: (2-3 sentences reasoning)
- Observed Pattern: (e.g., gather-scatter)

legitimate

v2_01197 — 2 txns

ground truth: not suspicious | 1,832 chars

You are an expert financial crime investigator reviewing patterns of financial activities and behaviors of involved accounts to identify potential cases of money laundering. The data is represented as a graph, where:
- Nodes are of type Account or Bank.
- Edges represent relationships of type transfers_to or belongs_to, and include metadata such as amount, currency, payment method, and timestamp.

For training purposes, you will be shown examples of subgraph typologies that are known to be either suspicious (indicative of laundering tactics) or non-suspicious (routine financial activity). These typologies illustrate common structural patterns in financial networks.

Task: Given a transaction (edge) with Transaction ID <ID>, along with its surrounding subgraph, determine whether the transaction is suspicious, or not suspicious. Your reasoning should be based on whether the surrounding subgraph resembles any of the suspicious typologies provided in the examples.
Test Example:
=== Transaction Subgraph (Case: v2_01197) ===

**Nodes:**
- acct_8056D5A70 (type: Account)
- acct_8056D6480 (type: Account)
- bank_015055 (type: Bank)
- bank_031755 (type: Bank)

**Edges:**
- acct_8056D5A70 belongs_to bank_015055
- acct_8056D6480 belongs_to bank_031755
- acct_8056D6480 transfers_to acct_8056D5A70 amount: 143.60 Euro via: Cash timestamp: 2022/09/02 01:31
- acct_8056D6480 transfers_to acct_8056D5A70 amount: 143.60 Euro via: Cash timestamp: 2022/09/09 07:20

SUMMARY STATISTICS:
  Total transactions: 2
  Unique accounts: 2
  Total volume: 287.20
  Avg transaction: 143.60
  Time span: 2 units
  Max out-degree: 2 (node: 031755_8056D6480)
  Max in-degree: 2 (node: 015055_8056D5A70)
Answer Format:
- Conclusion: Suspicious or Not Suspicious
- Explanation: (2-3 sentences reasoning)
- Observed Pattern: (e.g., gather-scatter)

suspicious

v2_00903 — 7 txns

ground truth: illicit | medium size

You are an expert financial crime investigator reviewing patterns of financial activities and behaviors of involved accounts to identify potential cases of money laundering. The data is represented as a graph, where:
- Nodes are of type Account or Bank.
- Edges represent relationships of type transfers_to or belongs_to, and include metadata such as amount, currency, payment method, and timestamp.

For training purposes, you will be shown examples of subgraph typologies that are known to be either suspicious (indicative of laundering tactics) or non-suspicious (routine financial activity). These typologies illustrate common structural patterns in financial networks.

Task: Given a transaction (edge) with Transaction ID <ID>, along with its surrounding subgraph, determine whether the transaction is suspicious, or not suspicious. Your reasoning should be based on whether the surrounding subgraph resembles any of the suspicious typologies provided in the examples.
Test Example:
=== Transaction Subgraph (Case: v2_00903) ===

**Nodes:**
- acct_80AF70A50 (type: Account)
- acct_80AF710E0 (type: Account)
- acct_80AF71220 (type: Account)
- acct_80AF711D0 (type: Account)
- bank_0028663 (type: Bank)
- bank_0128557 (type: Bank)
- bank_0128636 (type: Bank)
- bank_0228009 (type: Bank)

**Edges:**
- acct_80AF70A50 belongs_to bank_0028663
- acct_80AF710E0 belongs_to bank_0128557
- acct_80AF71220 belongs_to bank_0128636
- acct_80AF711D0 belongs_to bank_0228009
- acct_80AF71220 transfers_to acct_80AF71220 amount: 3415.92 Australian Dollar via: Reinvestment timestamp: 2022/09/01 00:11
- acct_80AF71220 transfers_to acct_80AF71220 amount: 30.32 Australian Dollar via: Reinvestment timestamp: 2022/09/01 00:13
- acct_80AF711D0 transfers_to acct_80AF711D0 amount: 20.41 Australian Dollar via: Reinvestment timestamp: 2022/09/01 15:07
- acct_80AF70A50 transfers_to acct_80AF70A50 amount: 13.11 Australian Dollar via: Reinvestment timestamp: 2022/09/01 16:47
- acct_80AF710E0 transfers_to acct_80AF711D0 amount: 2585.38 Australian Dollar via: ACH timestamp: 2022/09/09 03:46
- acct_80AF71220 transfers_to acct_80AF711D0 amount: 1751.92 Australian Dollar via: ACH timestamp: 2022/09/09 03:56
- acct_80AF70A50 transfers_to acct_80AF711D0 amount: 6851.05 Australian Dollar via: ACH timestamp: 2022/09/09 03:57

SUMMARY STATISTICS:
  Total transactions: 7
  Unique accounts: 4
  Total volume: 14,668.11
  Avg transaction: 2,095.44
  Time span: 7 units
  Max out-degree: 3 (node: 0128636_80AF71220)
  Max in-degree: 4 (node: 0228009_80AF711D0)
Answer Format:
- Conclusion: Suspicious or Not Suspicious
- Explanation: (2-3 sentences reasoning)
- Observed Pattern: (e.g., gather-scatter)

DATA

HI-Small FS — 3,753 cases (16 files)

system_prompt + 16 user_prompt files + manifest | ~264 MB total

Pre-processed prompts for enterprise Gemini 2.5 Pro. Download these files to your work machine, then run the loop below.

system_prompt.txt (122 KB) user_prompts_00.jsonl (38 MB, 500 cases) manifest.json (425 KB)

Usage on work machine (Vertex AI notebook):

import json, time, glob
import vertexai
from vertexai.preview.generative_models import GenerativeModel

vertexai.init(project="YOUR_PROJECT", location="us-central1")
model = GenerativeModel("gemini-2.5-pro")
params = {"temperature": 1.0, "max_output_tokens": 4096, "top_p": 0.95, "top_k": 1}

# Load
with open("system_prompt.txt", "r") as f:
    system_prompt = f.read()

user_prompts = []
for fname in sorted(glob.glob("user_prompts_*.jsonl")):
    with open(fname, "r") as f:
        for line in f:
            user_prompts.append(json.loads(line))

# Run
results = []
for i, up in enumerate(user_prompts):
    t0 = time.time()
    resp = model.generate_content(system_prompt + up["prompt"], generation_config=params, stream=False)
    ms = (time.time() - t0) * 1000
    u = resp.usage_metadata
    results.append({"case_id": up["case_id"], "response": resp.text,
                     "input_tok": u.prompt_token_count, "output_tok": u.candidates_token_count,
                     "total_tok": u.total_token_count, "ms": ms})
    if (i+1) % 50 == 0: print(f"  {i+1}/{len(user_prompts)}")

with open("gemini25pro_results.json", "w") as f:
    json.dump(results, f, indent=2, ensure_ascii=False)
print(f"Done! {len(results)} predictions saved")