What is Seerie v2?

The September 2026 release

Seerie v2 is the previous open GGUF of the on-device India fraud assistant. It is fine-tuned from Qwen3-1.7B on dataset v224 (36,669 conversations) and covers 22 core fraud types: UPI, OTP and KYC, Aadhaar misuse, digital arrest, phishing, and more. It runs fully offline.

Weights live at seerror-technologies/seerie-v224-GGUF. For new work, use Seerie v3 (seerie-v349-GGUF): three times the data, 2,954 extra scam topics, and a 100-prompt hand audit.

What Seerie v2 knows

22 core India fraud types

UPI fraud patternsCollect request bait, refund traps, and what to do next
Phishing SMS and callsFake KYC, courier duty, utility cut-off scripts
Digital arrest tacticsVideo call pressure, fake police and CBI scripts
Aadhaar safetyLock, unlock, masking, and misuse prevention
Indian cyber law basicsIT Act and DPDP Act, stated plainly
Where to reportcybercrime.gov.in and the 1930 helpline
English, Hinglish, 9 scriptsHindi, Bengali, Tamil, Telugu, Kannada, Malayalam, Gujarati, Punjabi, Odia
On device by designNo cloud inference step for the open weights
Example

A Hinglish exchange

Sample · UPI loss
User
Mera UPI se paise kat gaye, kya karoon?
Seerie
Sabse pehle turant apne bank ko call karke card/UPI block karayein.
Phir cybercrime.gov.in par complaint file karein ya 1930 helpline par call karein,
jitni jaldi report karenge, paise wapas milne ke chances utne zyada honge.
Apna transaction ID aur bank statement screenshot ready rakhein complaint ke liye.
Coverage

36,669 conversations

Responses point people to official channels such as the 1930 National Cyber Crime Helpline or cybercrime.gov.in.

TopicConversations
OTP / KYC Fraud10,416
UPI / Payment Fraud8,906
Aadhaar Misuse4,671
SIM Swap / Telecom Fraud2,324
Phishing (SMS, Email, Call)2,102
General Cyber Safety / Other1,904
IT Act / Cyber Law Guidance1,056
Digital Arrest Scam730
Lottery / Prize Scam639
Investment / Trading Scam638
Ransomware / Malware612
Loan App Fraud585
Sextortion / Blackmail486
Deepfake / AI Voice Cloning Scam384
Social Media Account Hacking329
OLX / Marketplace Scam279
Matrimonial / Romance Scam193
Electricity / Utility Bill Scam141
Data Privacy / DPDP Act116
Job / Work From Home Scam66
Cyberbullying / Harassment48
Army / Defence Impersonation Scam44
Total36,669
Model details

Fine tuned from Qwen3 1.7B

Causal language model, instruction-tuned for conversation. Fine-tuned from Qwen/Qwen3-1.7B with QLoRA and Unsloth, 3 epochs, on 36,669 curated conversations (dataset v224). License: Apache 2.0.

Quantizations

Available GGUF files

QuantizationSizeNotes
Q4_K_M Recommended1.11 GBDefault. Phones and low-memory devices.
Q6_K1.42 GBHigher-end phones and laptops.
F163.45 GBGPU machines, evaluation, or further quantization.

Get the files from huggingface.co/seerror-technologies/seerie-v224-GGUF.

How to run

LM Studio, Ollama, llama.cpp

Ollama

ollama run hf.co/seerror-technologies/seerie-v224-GGUF:Q4_K_M

llama.cpp

llama-cli -hf seerror-technologies/seerie-v224-GGUF:Q4_K_M

Or with a local Q4_K_M file from that repo:

llama-cli -m seerie-v224-q4_k_m.gguf -p "Mera UPI se paise kat gaye, kya karoon?"

LM Studio

  1. Search for seerror-technologies/seerie-v224-GGUF
  2. Load the model and leave the system prompt empty
  3. Use the chat template embedded in the GGUF file
  4. Set temperature to 0.3 to 0.4
Evaluation

58 / 70 safe on a manual reading

Seerie v2 was evaluated on 70 prompts in 14 categories. The automatic keyword harness used during training passed 63/70 (90%). A manual reading of every answer found 58 safe (82.9%) and 12 wrong. The later v3 paper compared both models on those 70 prompts: v3 gave 59 safe answers, v2 58. Breadth is what changed in v3, not a large drop in mistakes.

Full v224 audit: research page and Zenodo 10.5281/zenodo.23110076.

Changelog

How v2 sits in the line

v3 (October 2026) — current. Dataset v349, 107,575 conversations. Open the v3 model page.

v2 (September 2026) — this page. Dataset v224: 36,669 conversations; 9 Indian languages in native scripts; Q4_K_M, Q6_K and F16.

v1. Initial public release, trained on about 25,000 conversations in English, Hindi, and Hinglish.

Limitations

What Seerie v2 is not

  • Not a substitute for filing an official complaint. Urgent cases: 1930 or cybercrime.gov.in.
  • It can invent law sections, phone numbers, and websites. Check official sources.
  • Coverage stops at the 22 core topics in v224. Newer scam types are in v3.
  • Not legal, financial, or medical advice, and not a government product.
Links
Hugging Face · v2 GGUF
seerie-v224-GGUF · 298 downloads
Seerie v3 model page
current · dataset v349
Hugging Face · v3 GGUF
seerie-v349-GGUF
v224 research paper
manual audit · open PDF
Downloads & install
All official installs
Google Play
Install Seerie
Citation

How to cite

@misc{tiwari2026seerie,
  author       = {Tiwari, Jay},
  title        = {Seerie: An Offline 1.7B-Parameter Assistant for Scam Prevention in India,
                  and a Manual Audit of Its Failures},
  year         = {2026},
  note         = {Version 224},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.23110076},
  url          = {https://doi.org/10.5281/zenodo.23110076}
}
Previous release · v2 · dataset v224 · Apache 2.0
Get v2 on Hugging Face Open current v3