An offline assistant for Indian digital fraud
Seerie is a compact, on-device language model that helps people in India recognize, respond to, and report digital fraud, including UPI scams, OTP and KYC fraud, digital arrest, fake jobs and notices, fake donation appeals, online “friends” who ask for money, and about 4,400 other scam topics. It runs fully offline, so user questions and personal data never leave the device.
This page points to GGUF quantizations of Seerie v4 for llama.cpp, Ollama, and LM Studio: seerror-technologies/seerie-v407-GGUF. Previous models: seerie-v349-GGUF (v3) and seerie-v224-GGUF (v2). The Seerie app is live on Google Play. Install links also sit on the Downloads & install page.
On our tests v4 is about as accurate as v3, not more accurate. Every answer of this model on a 100-prompt test suite and on 84 new hand-written prompts (three answers each) was read and rated by an AI assistant (Claude). The ratings were not re-rated by a human. The full evaluation, including every rated answer, is published in Cleaner Data, the Same Error Rate (Zenodo 10.5281/zenodo.23731473).
India first safety knowledge
Trained to answer the questions people actually ask after a scare, in the language they use.
A real answer from the evaluation
Rated correct.
1) Shelter or rescue groups are usually registered on NGO Darpan (ngodarpan.gov.in), so verify their registration there.
2) If you can, visit the shelter yourself and see where they actually keep the animals.
3) A genuine group has official contact, a place, and often photographs of the work.
4) Money should go only to an account in the organisation's name — never an individual's personal UPI.
Did you send any money?
Helpline: 1930 | cybercrime.gov.in
Cleaner data, native Hindi, same error rate
| v3 (dataset v349) | v4 (dataset v407) | |
|---|---|---|
| Training conversations | 107,575 | 140,813 |
| Scam topics | ~3,000 | ~4,400 |
| Native Hindi (Devanagari) | early data only | new Hindi conversations in every batch since v401 |
| Law-section citations in training | 1,303 old records still cited sections | none: every citation removed or the record left out |
| Helpline numbers and websites | new data checked | all data checked; unverified numbers or unreachable government sites left out |
| Corrupted records | present | 1,319 left out, 866 turns repaired |
| Maximum length | 2,048 tokens | 3,072 tokens |
| Evaluation | 100-prompt suite, rated by hand | same suite, plus 84 hand-written prompts in English, Hinglish and Hindi, 3 answers each, plus an automatic fact checker |
140,813 conversations across India fraud types
Seerie v4 was trained on 140,813 conversations (dataset v407), about 4,400 scam topics. 45 scam topics (1,128 conversations) were held out for testing. Responses point people to official channels such as the 1930 National Cyber Crime Helpline or cybercrime.gov.in. The topic breakdown for the v3 corpus is on the v3 model page.
Fine tuned from Qwen3 1.7B
Seerie is a causal language model, instruction-tuned for conversation. Developed by Seerror Technologies (Jay Tiwari). Fine-tuned from Qwen/Qwen3-1.7B (Apache 2.0) with QLoRA and Unsloth, rank 32 / alpha 64, 3 epochs (26,460 steps), max length 3,072 tokens, loss on assistant replies only. Training data: 140,813 conversations (dataset v407). Evaluation: 86/100 safe answers (base model: 23/100); 76% safe on 84 new hand-written prompts. Languages: English, Hinglish, native Hindi, and 8 further Indian languages. License: Apache 2.0.
Available GGUF files
| File | Size | Notes |
|---|---|---|
seerie-v407-q4_k_m.gguf Recommended |
1.11 GB | Q4_K_M. Default. Phones and low-memory devices. |
Get the file from huggingface.co/seerror-technologies/seerie-v407-GGUF. Previous releases: seerie-v349-GGUF and seerie-v224-GGUF.
LM Studio, Ollama, llama.cpp
Ollama
ollama run hf.co/seerror-technologies/seerie-v407-GGUF:Q4_K_M
llama.cpp
llama-cli -hf seerror-technologies/seerie-v407-GGUF:Q4_K_M
Or with a local file:
llama-cli -m seerie-v407-q4_k_m.gguf -p "Mera UPI se paise kat gaye, kya karoon?"
LM Studio
- Search for
seerror-technologies/seerie-v407-GGUF, or download a.gguffile from Hugging Face - Load the model and leave the system prompt empty
- Use the chat template embedded in the GGUF file
- Set temperature to 0.3 to 0.4
Recommended settings
| Setting | Value |
|---|---|
| Temperature | 0.3 to 0.4 |
| System prompt | Leave empty (behaviour is trained into the model) |
| Chat template | Use the template embedded in the GGUF file (thinking disabled) |
86 / 100 safe on the previous suite
Seerie v4 and the untuned base model each answered the 100-prompt suite of the v3 report (35 English, 65 Hinglish, 20 categories). Every one of the 200 answers was read and rated correct, minor flaw, or wrong by an AI assistant. A human did not re-rate them.
| Seerie v4 | Seerie v3 | Base Qwen3-1.7B | |
|---|---|---|---|
| Safe answers (correct or minor flaw), 95% CI | 86 / 100 (78–92) | 87 / 100 | 23 / 100 (16–32) |
| Wrong answers | 14 | 13 | 77 |
| Wrong answers, English / Hinglish prompts | 5 of 35 / 9 of 65 | 2 of 35 / 11 of 65 | 12 of 35 / 65 of 65 |
| Safe on the 70 prompts every Seerie model answered | 57 | 59 | — |
84 new hand-written prompts
Written for this evaluation, outside the training templates: 56 scams (the model should warn), 17 genuine requests (it should help without alarm) and 11 requests to commit fraud or harm (it should refuse). Each was answered three times (252 answers).
| Safe answers | |
|---|---|
| All | 76% (192 / 252) |
| English / Hinglish / Hindi | 83% / 81% / 52% |
| Scam (warn) / Genuine (help) / Harmful (refuse) | 79% / 65% / 79% |
Fine-tuning helps a great deal, above all in Hinglish, where the base model failed every prompt. v4 is about as accurate as v3 and v2 on the original questions: on the same 70 prompts, v2, v3 and v4 give 58, 59 and 57 safe answers. Cleaning the data removed specific errors but not the error rate.
It still invents facts. Asked which IT Act section applies to a cybercrime FIR, it listed sections that do not exist, although no training conversation cites any section. It is too suspicious of genuine requests: it called a normal login OTP that the user requested “not normal”, and advised against paying by QR at a petrol pump. It refused every English request to write scam content, but gave safety tips to someone describing money-mule activity, and some Hindi harmful requests got weak or no refusals. On 45 held-out topics it passed 408 of 408 automatic checks, but 271 of those answers copy a training reply word for word.
Training loss and held-out loss
Seerie v4 was fine-tuned from Qwen3-1.7B (4-bit NF4 build from Unsloth) with QLoRA for 3 epochs on 140,813 conversations, on a single NVIDIA T4 GPU over several Kaggle sessions, with a peak learning rate of 2e-4, cosine decay and 8-bit AdamW. Final training loss 0.062; held-out loss 0.152 (measured at the end only). The adapter was merged and converted to GGUF with llama.cpp.
What changed in v4
v4 (October 2026). Retrained on dataset v407: 140,813 conversations, up from 107,575; about 4,400 scam topics; native Hindi. All law-section citations removed from training; unverified numbers, unreachable government websites and corrupted records left out. First release tested on hand-written prompts in English, Hinglish and Hindi, with an automatic fact checker. Weights: seerie-v407-GGUF.
v3 (October 2026). Dataset v349: 107,575 conversations; 2,954 new scam topics; first release rated by hand against the base model. Open the v3 model page or download seerie-v349-GGUF.
v2 (September 2026). Dataset v224: 36,669 conversations; 9 Indian languages in native scripts; Q4_K_M, Q6_K and F16. Open the v2 model page or download seerie-v224-GGUF.
v1. Initial public release, trained on approximately 25,000 conversations in English, Hindi, and Hinglish.
What Seerie is not
- Not a substitute for an official complaint. Urgent cases should always be reported to 1930 or cybercrime.gov.in.
- About 1 in 7 answers on the test suite and 1 in 4 on the hand-written test was wrong.
- It can state invented law sections, helpline numbers and websites with confidence. Verify such details with official sources.
- It sometimes treats genuine bank procedures as scams.
- Native Hindi answers are weaker than English and Hinglish answers.
- Many answers follow fixed templates, and the opening line does not always fit the question.
- The ratings were made by an AI assistant (Claude), not re-rated by a human. A full manual review is planned for the next model.
- Not suitable for replacing official complaints, providing legal, financial, or medical advice, or making automated decisions about individuals or transactions.
Where to go next
How to cite
@misc{tiwari2026seerie_v4,
title = {Cleaner Data, the Same Error Rate: A Third Answer-by-Answer Evaluation of Seerie, an Offline
Scam-Prevention Assistant for India, with a Hand-Written Test in English, Hinglish and Hindi},
author = {Tiwari, Jay},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.23731473},
url = {https://doi.org/10.5281/zenodo.23731473}
}