Osprey Narrator takes raw JSON output from the Osprey transaction monitoring engine and produces structured, analyst-ready compliance narratives — replacing hours of manual report writing with seconds of inference.

Model Details

Property	Value
Base Model	Qwen3-4B-Instruct-2507 (4-bit quantized)
Method	LoRA (Low-Rank Adaptation)
Framework	Unsloth + HuggingFace TRL
LoRA Rank	8
LoRA Alpha	16
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Trainable Params	16.5M / 4.0B (0.41%)
Max Sequence Length	4096 tokens

Performance

Metric	Value	Target
Test Perplexity	2.65	< 5.0
Test Loss	0.976	—
Final Train Loss	0.538	—
Validation Loss	0.538	—

Training Details

Parameter	Value
Training Examples	3,000
Validation Examples	375
Test Examples	375
Epochs	1
Batch Size	2 (effective 8 with gradient accumulation)
Learning Rate	2e-5
Warmup Steps	20
Total Steps	375
Training Time	~70 min on NVIDIA T4 (Google Colab)
Precision	FP16
GPU	NVIDIA T4 16GB

Loss Curve

Step	Train Loss	Validation Loss
100	0.640	0.624
200	0.560	0.553
300	0.539	0.538

Steady decrease with no divergence — healthy training with no overfitting.

What It Does

Given a JSON evaluation from the Osprey transaction monitoring engine, the model generates a structured compliance narrative with these sections:

Alert Summary — Alert ID, transaction type, amount, parties, decision
Transaction Details — Contextual description of the transaction
Risk Assessment — Overall risk analysis with score interpretation
Rules Triggered — Table of all 12 FATF rules with scores
Typology Analysis — Table of 6 FATF typologies with match status
Narrative — Human-readable explanation connecting the dots
Recommended Actions — Next steps for the compliance team

FATF Rules Covered

Rule ID	Name
FATF-R001	High-Value Transaction
FATF-R002	Structuring Detection
FATF-R003	Rapid Movement of Funds
FATF-R004	Geographic Risk
FATF-R005	Unusual Transaction Pattern
FATF-R006	Shell Company Indicator
FATF-R007	PEP Transaction
FATF-R008	Round Amount Transaction
FATF-R009	Cross-Border Wire Transfer
FATF-R010	Dormant Account Activity
FATF-R011	Currency Exchange Anomaly
FATF-R012	Third-Party Payment

Typologies Covered

ID	Name
FATF-T001	Structuring / Smurfing
FATF-T002	Trade-Based Money Laundering
FATF-T003	Shell Company Layering
FATF-T004	PEP Corruption Proceeds
FATF-T005	Funnel Account Activity
FATF-T006	Currency Exchange Laundering

Quick Start

With Unsloth (recommended for inference)

python

from unsloth import FastLanguageModel
import json

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="josephgoksu/osprey-narrator-v0.1",
    max_seq_length=4096,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

# Your Osprey evaluation JSON
evaluation = {
    "decision": "ALRT",
    "composite_score": 724.50,
    "threshold": 500.0,
    "transaction": {
        "type": "wire_transfer",
        "amount": 147500.00,
        "currency": "USD",
        "originator": {"name": "Atlas Global Ventures", "country": "AE"},
        "beneficiary": {"name": "Pinnacle Commodities Ltd.", "country": "NG"}
    },
    # ... full evaluation JSON
}

messages = [
    {"role": "system", "content": "You are an experienced AML/CFT compliance analyst..."},
    {"role": "user", "content": json.dumps(evaluation)}
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    enable_thinking=False,
    return_tensors="pt",
).to("cuda")

outputs = model.generate(
    input_ids=inputs,
    max_new_tokens=1024,
    temperature=0.7,
    top_p=0.9,
)

narrative = tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True)
print(narrative)

With HuggingFace Transformers + PEFT

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import json

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-4B-Instruct-2507",
    device_map="auto",
    load_in_4bit=True,
)
tokenizer = AutoTokenizer.from_pretrained("josephgoksu/osprey-narrator-v0.1")

# Apply LoRA adapter
model = PeftModel.from_pretrained(base_model, "josephgoksu/osprey-narrator-v0.1")

# Generate (same message format as above)
messages = [
    {"role": "system", "content": "You are an experienced AML/CFT compliance analyst..."},
    {"role": "user", "content": json.dumps(evaluation)}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")
outputs = model.generate(inputs, max_new_tokens=1024, temperature=0.7)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

With MLX on Apple Silicon (macOS)

pip install mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load(
    "josephgoksu/osprey-narrator-v0.1",
    adapter_path="josephgoksu/osprey-narrator-v0.1",
)
# Use the same chat template approach

Example Output

Input: ALRT evaluation with score 724.50, wire transfer of $147,500 USD from AE → NG

Output (abbreviated):

Alert Summary

Alert ID: eval-7a3b2c1d-4e5f-6a7b-8c9d-0e1f2a3b4c5d

Transaction Type: wire_transfer

Amount: 147,500.00 USD

Decision: ALRT (Alert Generated)

Composite Score: 724.50 (threshold: 500.0)

Transaction Details

A wire transfer of 147,500.00 USD was initiated by Atlas Global Ventures (AE) to Pinnacle Commodities Ltd. (NG). The transaction involves a cross-border movement between high-risk jurisdictions...

Risk Assessment

The composite risk score of 724.50 significantly exceeds the alert threshold of 500.0, indicating elevated risk. Seven of twelve FATF rules were triggered, with High-Value Transaction (0.875) and Geographic Risk (0.810) contributing the highest individual scores...

Rules Triggered

Rule Score Triggered
FATF-R001 High-Value Transaction 0.875 Yes
FATF-R003 Rapid Movement of Funds 0.720 Yes
FATF-R004 Geographic Risk 0.810 Yes
... ... ...

Recommended Actions

Escalate to senior compliance officer for SAR filing review

Request enhanced due diligence on both counterparties

Investigate related transactions within the past 90 days

Training Data

The model was trained on synthetic data only — no real customer or transaction data was used. The synthetic dataset was generated to cover:

5 scenario types: Clean transactions (40%), single rule triggers (15%), multi-rule below threshold (10%), typology matches (20%), critical alerts (15%)
Decision distribution: ~35% ALRT (alert) / ~65% NALT (no alert)
12 FATF rules and 6 typologies with realistic score distributions
5-8 phrasing variants per rule/typology to prevent memorization

Limitations

Synthetic training only — may not capture all real-world edge cases
v0.1 POC — not yet validated against production compliance requirements
English only — narratives are generated in English
Token limit — very complex evaluations may be truncated at 4096 tokens
No real regulatory validation — output should be reviewed by compliance professionals before use in actual SAR filings

Architecture

Osprey TM Engine → JSON Evaluation → Osprey Narrator → SAR Narrative
                                      (this model)

The model sits downstream of the Osprey transaction monitoring engine. It receives the structured evaluation output (rules, typologies, scores, decision) and transforms it into human-readable compliance narratives.

License

Apache 2.0

Citation

bibtex

@misc{osprey-narrator-2026,
  title={Osprey Narrator: Fine-Tuned SAR Narrative Generation},
  author={Joseph Goksu},
  year={2026},
  url={https://huggingface.co/josephgoksu/osprey-narrator-v0.1}
}

Rule	Score	Triggered
FATF-R001 High-Value Transaction	0.875	Yes
FATF-R003 Rapid Movement of Funds	0.720	Yes
FATF-R004 Geographic Risk	0.810	Yes
...	...	...