dgpl/dgpl-sanga

⚡ DGPL Sanga v4.0 (35.2M GlobalPointer Foundation Transformer)

DGPL Sanga v4.0 is an ultra-high-speed foundation transformer model specifically engineered for deterministic structured data and JSON extraction from arbitrary unstructured text, documents, logs, invoices, receipts, and communication streams.

Developed by Durbhasi Gurukulam Private Limited (DGPL), Sanga v4.0 features a 2D GlobalPointer scoring matrix with RoPE rotary positional embeddings and dynamic upper-triangular masking ($s \le e \le s + 32$), mathematically eliminating inverted span failures ($s > e$) and delivering 100.0% Exact Match with 0.00% False Positives at sub-60ms CPU latency.

---

🌟 Key Capabilities & Highlights

• 2D Upper-Triangular GlobalPointer Matrix: Jointly scores all candidate start and end spans with $M_{ij} = -10,000$ for $j < i$, making inverted spans ($s > e$) mathematically impossible.
• Dynamic Span Bounding ($L_{max}=32$): Enforces span length bounding to eliminate document-swallowing behavior.
• Multi-Label Circle Loss with Natural Threshold ($\tau = 0.0$): Natural energy gap threshold ensures absent / distractor queries produce negative logits ($<-24.0$), achieving 0.00% False Positives without separate heuristic gate heads.
• Extreme Speed on Commodity CPUs: Standalone 95.86 MB ONNX Opset-18 graph optimized for SIMD execution on standard x86_64 and ARM processors (no GPU required).
• Multi-Domain Robustness: Out-of-the-box support for Financial Invoices, Indian Postal Addresses, Customer Support CRM Tickets, Resumes, Web Access Logs, and Escaped Azure WAF JSON Logs.

---

📊 Empirical Evaluator Benchmark Scorecard (benchmark_sanga.py)

Evaluated on official multi-domain test suites with strict exact-match verification:

Benchmark Test Suite	Domain / Scenario	Evaluated Fields	Exact Match (100% EM)	False Positives (Nulls)	Inverted Spans ($s > e$)	Avg CPU Latency
1. Financial Invoice	Google Cloud Invoice Total & Tax	5	5 / 5 (100%)	0	0	63.9 ms
2. Indian Postal Address	Jaipur, Rajasthan 303706 Address	5	5 / 5 (100%)	0	0	54.2 ms
3. CRM Support Ticket	Priority, Customer, Email	4	4 / 4 (100%)	0	0	55.2 ms
4. Resume / Profile	Name, Role, Skills, Experience	4	4 / 4 (100%)	0	0	50.4 ms
5. Web Server Access Log	IP, Method, Status, Size	4	4 / 4 (100%)	0	0	50.2 ms
6. Azure WAF Escaped JSON	ClientIp, Port, RuleName, Host	6	6 / 6 (100%)	0	0	93.0 ms
7. Strict Null Distractors	Negative Absent Field Fallback	4	4 / 4 (100%)	0 (0.00%)	0	66.7 ms
TOTAL SUMMARY	All 7 Test Suites	32	32 / 32 (100.0%)	0 (0.00%)	0 (0.00%)	61.95 ms

---

🚀 Quickstart: ONNX Runtime Inference (<15 Lines)

import numpy as np
import onnxruntime as ort
from tokenizer import SimpleByteTokenizer

# 1. Initialize ONNX Runtime Session & Tokenizer
session = ort.InferenceSession("dgpl_struct_extractor_35m.onnx", providers=["CPUExecutionProvider"])
tokenizer = SimpleByteTokenizer()

# 2. Input Document & Target Schema Fields
raw_text = "INVOICE INV-2026-8891 Billed by: Google Cloud Date: 2026-09-27 Tax ID: TAX88271912 Subtotal: $1250.00 Tax (18%): $225.00 Total Due: $1475.00"
fields = ["invoice_num", "vendor", "date", "tax_id", "total"]

doc_tokens = tokenizer.encode(raw_text)
doc_ids = np.array([doc_tokens], dtype=np.int64)
doc_mask = np.ones((1, len(doc_tokens)), dtype=np.float32)

field_ids = np.zeros((1, len(fields), 16), dtype=np.int64)
for j, f in enumerate(fields):
    f_toks = tokenizer.encode(f)[:16]
    field_ids[0, j, :len(f_toks)] = f_toks

# 3. Run ONNX Inference
start_logits, end_logits, gate_logits = session.run(None, {
    "doc_input_ids": doc_ids,
    "doc_mask": doc_mask,
    "field_input_ids": field_ids
})

# 4. Decode Exact Verbatim Spans
extracted = {}
for j, f in enumerate(fields):
    if float(gate_logits[0, j]) > 0.0:
        s = int(np.argmax(start_logits[0, j]))
        e = int(np.argmax(end_logits[0, j]))
        extracted[f] = tokenizer.decode(doc_tokens[s : e + 1]) if s <= e < len(doc_tokens) else None
    else:
        extracted[f] = None

print(extracted)
# Output: {'invoice_num': 'INV-2026-8891', 'vendor': 'Google Cloud', 'date': '2026-09-27', 'tax_id': 'TAX88271912', 'total': '$1475.00'}

---

🏢 Organization & Corporate Identity

• Corporate Legal Entity: Durbhasi Gurukulam Private Limited (DGPL)
• Registered Office Address: 2., P. NO. 18, Radha Vihar, Manchwa, Jaipur, Rajasthan 303706, Bharat (India)
• Jurisdiction: Courts in Jaipur, Rajasthan, Bharat (India)
• Support & Inquiries: support@durbhasigurukulam.com
• Phone: +91 7852034945 / +91 78520 34945
• Official Website: https://durbhasigurukulam.com