naazimsnh02/fraud-financial-crime-qwen3-sft-v2

Fraud Detection & Financial Crime Intelligence — Conversational SFT Dataset

A conversational (ChatML) supervised fine-tuning dataset for training an LLM (built and tuned for Qwen3-14B) to act as an enterprise fraud-detection and financial-crime investigation assistant. Every example teaches the model to deliver real-time risk scoring, explainable alerts, and recommended next actions with human-in-the-loop (HITL) oversight — across both card fraud and anti-money-laundering (AML).

Trained model: naazimsnh02/fraudsentinel-qwen3-14b-lora · naazimsnh02/fraudsentinel-qwen3-14b-merged

---

Capabilities the model learns

1. Real-time risk scoring — a calibrated score (0–100 / 0.0–1.0) with a LOW / MEDIUM / HIGH / CRITICAL band.
2. Explainable alerts — specific, evidence-grounded red flags drawn from the actual transaction, not generic boilerplate.
3. Typology classification — names the most likely fraud / laundering pattern (primary + secondary).
4. Recommended next action — a concrete decision on a 6-level investigator taxonomy.
5. SAR support — flags when a Suspicious Activity Report is warranted, with a one-line rationale.
6. Human-in-the-loop dialogue — multi-turn investigator follow-ups ("Why this risk level?", "What else should I check?").

---

Two complementary answer styles

Both styles are grounded in the same real, engineered features.

1. Structured JSON (task = structured)

Dashboard- and pipeline-ready output:

{
  "risk_score": 0.84,
  "risk_level": "HIGH",
  "conclusion": "FRAUDULENT",
  "primary_typology": "card-not-present account takeover / stolen-card online cash-out",
  "secondary_typology": "account_takeover",
  "key_signals": ["amount_exceeds_category_p95", "high_risk_merchant_category", "unusual_hour_activity"],
  "explanation": "Transaction amount $828.62 exceeds the 95th-percentile ($383.21) for 'misc_net' purchases; ...",
  "feature_importance": {"amount_exceeds_category_p95": 0.46, "high_risk_merchant_category": 0.28, "unusual_hour_activity": 0.26},
  "recommended_action": "AUTO_BLOCK",
  "sar_required": false,
  "sar_rationale": null
}

2. Analyst prose (task = explain | score | recommend | multiturn)

Investigator-facing narrative with risk score, red flags, typology, and recommended action — plus multi-turn HITL follow-ups ("Why this risk level?", "What else should I check?", "Customer confirmed legit — now what?").

6-level investigator action taxonomy (HITL)

AUTO_APPROVE → APPROVE_WITH_MONITORING → STEP_UP_AUTH → TEMPORARY_HOLD → AUTO_BLOCK → SAR_REVIEW

---

Data Provenance — Grounded Label Generation (No Hallucinated Labels)

The dataset is synthesized from two public, verified tabular sources using label-grounded generation: the assistant's conclusion is always derived from the ground-truth label, and every red flag is computed from engineered features of the actual transaction. No teacher LLM is used, so there are no hallucinated labels and no label leakage — the model learns to interpret raw transaction fields, not echo injected tags.

Domain	Source dataset	Real label used	Coverage
Card fraud (card-not-present)	pointe77/credit-card-transaction (Sparkov, 1.3M tx)	is_fraud	risk scoring, alerts, takeover / card-testing typologies
Money laundering	eexzzm/IBM-Transactions-for-Anti-Money-Laundering-HI-Small-Trans (IBM AML, 5M tx)	Is Laundering	AML typologies, SAR-oriented actions

Engineered signals → tagged red flags

Card transactions

• Amount vs. per-category 95th-percentile
• High-risk merchant channel (shopping_net / misc_net / grocery_pos)
• Off-hours / unusual-hour timing
• Cardholder → merchant geo distance (haversine)
• 24-hour velocity (transaction count + spend, time-since-last)
• Large absolute ticket size

AML inter-bank transfers

• Fan-out / fan-in / gather-scatter (account out-/in-degree)
• Self-loops / round-tripping
• Round-number amounts
• Cross-currency conversion
• ACH channel over-representation
• Rapid pass-through (mule) movement

feature_importance is real, not LLM-guessed

Each detection rule carries a fixed weight. The emitted importances are the normalized weights of the rules that actually fired for that transaction (they sum to 1.0). Fully deterministic and auditable.

False-positive control built into the labels

Legitimate cases deliberately suppress strong structural tags (a high-degree account can be a legitimate business; a large purchase can be legitimate spend). This keeps the score, conclusion, and action coherent and teaches the model not to over-flag — directly supporting false-positive reduction.

---

Composition

• 11,816 conversations (≈ 6,000 card + ≈ 6,000 AML)
• Class balance: ~44% fraud/suspicious, ~56% legitimate. Real fraud is <1%, so the set is deliberately balanced; negatives include hard negatives (high-amount or high-risk-category legitimate activity) so the model learns to say "Not suspicious."
• Task mix: structured (3.4K) · explain (3.4K) · recommend (1.7K) · multiturn (1.7K) · score (~1.6K)
• Splits: train (11,016 examples) and test (800 examples)

Columns

Column	Description
messages	ChatML conversation (system / user / assistant)
source	cc (card) or aml (money laundering)
label	Ground truth: 0 legitimate, 1 fraud/suspicious
task	structured / explain / score / recommend / multiturn

Evaluation note: the included test split is balanced for convenient evaluation. For a realistic production estimate, re-sample a held-out set at the natural class imbalance (<1% fraud) and report precision / recall, not accuracy.

---

Fine-Tuning Recipe (Qwen3-14B — Validated)

This is the exact configuration used to train naazimsnh02/fraudsentinel-qwen3-14b-lora, verified on AMD MI300X with Unsloth 2026.6.1:

from unsloth import FastLanguageModel
from trl import SFTTrainer, SFTConfig
from datasets import load_dataset
import torch

# Load base model with LoRA
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/Qwen3-14B",
    max_seq_length = 4096,
    dtype = torch.bfloat16,
    load_in_4bit = False,
)
model = FastLanguageModel.get_peft_model(
    model,
    r = 16,
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj"],
    lora_alpha = 32,
    lora_dropout = 0,
    bias = "none",
    use_gradient_checkpointing = "unsloth",
    random_state = 3407,
)

# Load and format dataset
ds = load_dataset("naazimsnh02/fraud-financial-crime-qwen3-sft-v2")

def apply_template(example):
    text = tokenizer.apply_chat_template(
        example["messages"],
        tokenize=False,
        add_generation_prompt=False,
        enable_thinking=False,   # thinking OFF during SFT
    )
    return {"text": text}

ds = ds.map(apply_template)

# Train
trainer = SFTTrainer(
    model     = model,
    tokenizer = tokenizer,
    train_dataset = ds["train"],
    args = SFTConfig(
        output_dir                    = "./fraudsentinel_checkpoints",
        save_strategy                 = "steps",
        save_steps                    = 100,
        save_total_limit              = 5,
        dataset_text_field            = "text",
        max_seq_length                = 4096,
        packing                       = False,
        per_device_train_batch_size   = 2,
        gradient_accumulation_steps   = 8,     # effective batch = 16
        num_train_epochs              = 2,
        warmup_ratio                  = 0.05,
        learning_rate                 = 1e-4,
        lr_scheduler_type             = "cosine",
        bf16                          = True,
        optim                         = "adamw_8bit",
        padding_free                  = True,
        weight_decay                  = 0.001,
        logging_steps                 = 10,
        seed                          = 3407,
        dataloader_num_workers        = 4,
    ),
)
trainer.train()

Actual training results (AMD MI300X, 192 GB VRAM):

Metric	Value
Total steps	1,378
Train loss	0.2467
Training time	70.5 min (4,230 s)
Peak VRAM	39.8 GB (20.8% of 192 GB)
LoRA VRAM overhead	12.0 GB (6.3% of max)

---

Intended Use & Limitations

• Prototype / research use. Source data is synthetic / semi-synthetic; the typologies and scoring heuristics are illustrative, not a production-validated risk model.
• Do not deploy for real customer adjudication without independent validation, bias review, and a human-in-the-loop control layer.
• Names, accounts, and card numbers originate from the synthetic source datasets; card numbers are masked to last-4.
• Feature importance values are deterministic heuristics from the data generation pipeline, not SHAP or model-derived explanations.

---

Citation

Source datasets: Sparkov (credit-card transaction generator) and IBM Transactions for Anti-Money-Laundering (Altman et al., NeurIPS 2023, arXiv:2306.16424). Method inspired by arXiv:2507.14785 (LLMs for AML graphs) and arXiv:2506.11635 (LLM-based credit-card fraud investigation).