Instructions to use Anbukumaran1/tnea-advisor with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Anbukumaran1/tnea-advisor with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Anbukumaran1/tnea-advisor")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Anbukumaran1/tnea-advisor")
model = AutoModelForCausalLM.from_pretrained("Anbukumaran1/tnea-advisor", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

PEFT
How to use Anbukumaran1/tnea-advisor with PEFT:
```
Task type is invalid.
```
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Anbukumaran1/tnea-advisor with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Anbukumaran1/tnea-advisor"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Anbukumaran1/tnea-advisor",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Anbukumaran1/tnea-advisor

SGLang

How to use Anbukumaran1/tnea-advisor with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Anbukumaran1/tnea-advisor" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Anbukumaran1/tnea-advisor",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Anbukumaran1/tnea-advisor" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Anbukumaran1/tnea-advisor",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Anbukumaran1/tnea-advisor with Docker Model Runner:
```
docker model run hf.co/Anbukumaran1/tnea-advisor
```

TNEA-Advisor

TNEA-Advisor is a LoRA fine-tuned and merged derivative of Qwen2.5-3B-Instruct, specialized for Tamil Nadu Engineering Admissions (TNEA).

The model is designed to estimate TNEA ranks and shortlist engineering colleges by cutoff tier, using historical category-wise cutoff data from the Directorate of Technical Education (DoTE), Tamil Nadu.

A 4-bit GGUF build is also available for lightweight, local, and offline inference through Ollama:

Ollama — anbukumaran/tnea-advisor

Important: TNEA-Advisor provides estimates based on historical data. It is not an official DoTE/TNEA system and should not be treated as a guarantee of admission.

✨ What It Does

TNEA-Advisor is a domain-specific language model built around the TNEA counselling use case.

Given information such as:

TNEA cutoff mark
Community category
Preferred engineering branch

the model can estimate:

State General Rank
Community Rank
Suitable colleges based on historical cutoff patterns
College recommendations organized into three tiers:

Tier	Meaning
🚀 REACH	More ambitious options with comparatively tighter historical cutoffs
🎯 TARGET	Strong, realistic options based on historical trends
🛡️ SAFE	Options with comparatively more favorable historical cutoff margins

Supported Community Categories

OC · BC · BCM · MBC · SC · SCA · ST

📌 Model Details

Property	Details
Developed by	AnbuKumaran A
Model type	Causal decoder-only transformer (Qwen2 architecture), LoRA fine-tuned and merged
Base model	Qwen/Qwen2.5-3B-Instruct
Parameters	3B
Precision	FP16 merged Safetensors
Quantized build	4-bit GGUF `q4_K_M` via Ollama
Tokenizer	Original Qwen2.5 BPE tokenizer, unmodified through the merge
Language	English
License	Qwen Research License

This repository contains the merged standalone model, rather than the original LoRA adapter.

🧩 LoRA Configuration

Hyperparameter	Value
Rank (`r`)	16
Alpha	32
Dropout	0.05
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
Learning rate	`2e-4`
Epochs	3
Optimizer	`adamw_torch`

After training, the LoRA adapter was merged into the base model using peft's:

merge_and_unload()

The resulting repository contains the merged model weights, not the bare LoRA adapter.

🎯 Intended Use

In Scope

TNEA-Advisor is intended for:

Cutoff-mark-based State General Rank estimation
Cutoff-mark-based Community Rank estimation
Historical TNEA cutoff analysis
College shortlisting based on historical cutoff tiers
TNEA counselling exploration across supported community categories

Out of Scope

TNEA-Advisor is not intended to replace:

Official DoTE/TNEA counselling systems
Official TNEA notifications and publications
Qualified admissions counsellors
General-purpose conversational AI outside its trained domain

Always cross-check important admission decisions against the latest official DoTE/TNEA information.

🚀 Getting Started

Using Hugging Face Transformers

Install the required dependencies:

pip install transformers torch

Load the model:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Anbukumaran1/tnea-counselling-qwen2.5-3b"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

messages = [
    {
        "role": "system",
        "content": "You are TNEA-Advisor, a data-driven TNEA admissions strategist."
    },
    {
        "role": "user",
        "content": "Cutoff: 198.50/200, Community: OC, Preferred Branch: CSE"
    }
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

output_ids = model.generate(
    input_ids,
    max_new_tokens=512,
    temperature=0.2,
    top_p=0.9,
    do_sample=True
)

response = tokenizer.decode(
    output_ids[0][input_ids.shape[-1]:],
    skip_special_tokens=True
)

print(response)

🦙 Ollama — 4-bit GGUF

For a lighter and more CPU-friendly local deployment:

ollama run anbukumaran/tnea-advisor

The Ollama build uses a 4-bit GGUF q4_K_M quantization for efficient local inference.

💡 Example Input

Cutoff: 198.50/200
Community: OC
Preferred Branch: CSE

The model can use these inputs to estimate ranks and organize suitable colleges into:

🚀 REACH
🎯 TARGET
🛡️ SAFE

📚 Training Data

TNEA-Advisor was fine-tuned on historical, category-wise TNEA cutoff records sourced directly from the Directorate of Technical Education (DoTE), Tamil Nadu.

Dataset Coverage

Attribute	Coverage
Academic years	2021–2025
Historical span	5 academic years
Data volume	100,000+ official data points
Community categories	OC, BC, BCM, MBC, SC, SCA, ST
Colleges	Participating engineering colleges
Programs	Engineering branches and historical cutoff records

The model is intentionally specialized around historical TNEA admissions data rather than general-purpose knowledge.

📊 Evaluation

TNEA-Advisor was evaluated against the base Qwen2.5-3B-Instruct model using an automated LLM-as-a-judge evaluation pipeline.

Evaluation Setup

Judge model: Gemini 2.5 Flash
Structured outputs: Pydantic
Evaluation scenarios: 50 real-world cutoff scenarios

Results

Metric	Base Qwen2.5-3B-Instruct	TNEA-Advisor	Δ
Accuracy	7.0 / 10	10.0 / 10	+42.9%
Hallucination Resistance	7.0 / 10	10.0 / 10	+42.9%
Relevance	8.0 / 10	10.0 / 10	+25.0%
Overall	7.3 / 10	10.0 / 10	+36.4%

Key Finding

The largest measured improvement was hallucination resistance.

Fine-tuning on grounded, category-wise cutoff data substantially reduced the base model's tendency to produce plausible-sounding but incorrect rank or college recommendations within the evaluated TNEA scenarios.

Evaluation note: These results come from a relatively small 50-scenario automated benchmark. They should be interpreted as an evaluation signal rather than a comprehensive measurement of real-world admission accuracy.

🏗️ Architecture

                    ┌─────────────────────────┐
                    │   Qwen2.5-3B-Instruct   │
                    └────────────┬────────────┘
                                 │
                                 ▼
                        LoRA Fine-Tuning
                                 │
                                 ▼
                    ┌─────────────────────────┐
                    │ TNEA Historical Cutoffs │
                    │       2021–2025         │
                    │      100,000+ points    │
                    └────────────┬────────────┘
                                 │
                                 ▼
                       LoRA Adapter Merge
                                 │
                                 ▼
                    ┌─────────────────────────┐
                    │      TNEA-Advisor       │
                    │    Merged FP16 Model    │
                    └────────────┬────────────┘
                                 │
                    ┌────────────┴────────────┐
                    ▼                         ▼
            Transformers / HF           4-bit GGUF
                    │                         │
                    ▼                         ▼
              GPU Inference                Ollama

💻 Deployment Options

Deployment	Format	Recommended Use
Hugging Face Transformers	FP16 Safetensors	GPU inference and development
Ollama	4-bit GGUF `q4_K_M`	Lightweight local inference
Offline deployment	Local model files	Offline experimentation and use

⚠️ Limitations & Biases

Domain-Narrow by Design

TNEA-Advisor is optimized specifically for TNEA cutoff estimation and college shortlisting. Its performance outside this domain is not verified.

Historical Data

TNEA cutoffs can change from year to year due to factors including:

Seat-matrix changes
New colleges or programs
Policy revisions
Changes in applicant preferences
Changes in competition

Historical predictions should therefore be treated as informed estimates, not guarantees.

Not an Official Source

This model does not replace official DoTE/TNEA counselling information, notifications, or professional admissions advice.

Evaluation Scope

The reported evaluation uses a 50-scenario LLM-judged benchmark. While useful as an automated signal, it has not been validated through:

Large-scale human review
A broad independently sourced test set
Real-world admission outcomes

Uneven Category and Branch Coverage

Performance may vary across community categories and engineering branches, particularly where historical training data is relatively sparse.

Language Limitation

The model is documented and evaluated primarily for English-language input.

Performance on Tamil and Tanglish/code-mixed queries has not been specifically evaluated.

Inherited Base-Model Limitations

General-purpose biases, weaknesses, and failure modes present in Qwen2.5-3B-Instruct may still appear in the fine-tuned model.

📜 License

TNEA-Advisor is a fine-tuned derivative of Qwen/Qwen2.5-3B-Instruct and is distributed under the same Qwen Research License Agreement, rather than Apache 2.0.

In summary:

The model is intended for research and non-commercial use unless the applicable Qwen commercial-use requirements are separately satisfied.
Redistribution of the merged or further fine-tuned weights carries the applicable license obligations forward.
Review the full Qwen Research License Agreement before using this model in a commercial product or service.

License notice: This section is a summary for convenience and is not a substitute for reading the full license.

👤 Model Card Contact

AnbuKumaran A

GitHub · LinkedIn

🙏 Acknowledgements

Built by AnbuKumaran A using the Qwen2.5-3B-Instruct base model and historical TNEA/DoTE data.

The project aims to make historical engineering admissions data more accessible and useful for students exploring college options across Tamil Nadu.

⚖️ Disclaimer

TNEA-Advisor is an experimental, data-driven admissions estimation model.

Its predictions are based on historical information and may be incorrect, especially when current-year conditions differ from historical trends.

Always verify important admission decisions using the latest official TNEA/DoTE publications and counselling information.

Downloads last month: -

Safetensors

Model size

3B params

Tensor type

F16

Model tree for Anbukumaran1/tnea-advisor

Base model

Qwen/Qwen2.5-3B

Finetuned

Qwen/Qwen2.5-3B-Instruct

Adapter

(1330)

this model