---
base_model: unsloth/gemma-2-9b
tags:
- text-generation-inference
- transformers
- unsloth
- gemma2
- trl
- local-finetuning
license: apache-2.0
language:
- en
- te
datasets:
- eswardivi/Telugu_InstructData
# widget:
# - text: "### Task\nGenerate a SQL query to answer the following question:\n`How many heads of the departments are older than 56?`\n\n### Database Schema\nThe query will run on a database with the following schema:\nCREATE TABLE head (age INTEGER)\n\n### Answer\nGiven the database schema, here is the SQL query that answers `How many heads of the departments are older than 56?`:\n```sql"
#   example_title: "One Table"
# - text: "### Task\nGenerate a SQL query to answer the following question:\n`How many departments are led by heads who are not mentioned?`\n\n### Database Schema\nThe query will run on a database with the following schema:\nCREATE TABLE management (department_id VARCHAR);\nCREATE TABLE department (department_id VARCHAR)\n\n### Answer\nGiven the database schema, here is the SQL query that answers `How many departments are led by heads who are not mentioned?`:\n```sql"
#   example_title: "Two Tables" 
---

# Thanks for being patient! 💜💜

# Model Card for Model ID

<!-- Provide a quick summary of what the model is/does. -->

A fine-tuned version of Gemma2 9B for the Telugu Language Instruct on `eswardivi/Telugu_InstructData` dataset using [Unsloth](https://github.com/unslothai/unsloth).
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)

## Model Details

### Model Description

<!-- Provide a longer summary of what this model is. -->
This model has been finetuned with `eswardivi/Telugu_InstructData` on `unsloth/gemma-2-9b`. This is to ensure there is a raise in Telugu Language Model usage in daily applications. This model will respond in english text where as the content is in telugu language. 

For example instead of `ఎలా ఉన్నావ్?` it will respond with `ela unnav?`. This is to ensure the model is not using more tokens for unicode characters.

- **Developed by:** pavankumarbalijepalli
- **Model type:** CASUAL_LM
- **Language(s) (NLP):** English, Telugu
- **License:** MIT
- **Finetuned from model:** [unsloth/gemma-2-9b](https://huggingface.co/unsloth/gemma-2-9b)

### Model Sources

<!-- Provide the basic links for the model. -->

- **Repository:** [HappyDevCorp/ai-ft-tellm](https://github.com/HappyDevCorp/ai-ft-tellm/blob/main/finetunes/Gemma2_(9B)_Alpaca.ipynb)

## Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

Model is supposed to be used for the cases where you have a natural language question in telugu language to get a response in return in telugu language. The context should be below 8192 tokens. The output will be generated in telugu.


### Downstream Use

<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
```python
# WITH UNSLOTH - RECOMMENDED
from unsloth import FastLanguageModel

max_seq_length = 2048  # Choose any! We auto support RoPE Scaling internally!
dtype = (
    None
)
load_in_4bit = True

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="pavankumarbalijepalli/telLM-gemma2-9b",  # YOUR MODEL YOU USED FOR TRAINING
    max_seq_length=max_seq_length,
    dtype=dtype,
    load_in_4bit=load_in_4bit,
)
FastLanguageModel.for_inference(model)  # Enable native 2x faster inference

prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{}

### Input:
{}

### Response:
{}"""

inputs = tokenizer(
    [
        prompt.format(
            "Naa peru enti?",  
            "Pavan Kumar ninnu aduguthunnadu",  
            "",  
        )
    ],
    return_tensors="pt",
).to("cuda")

outputs = model.generate(**inputs, max_new_tokens=64, use_cache=True)
tokenizer.batch_decode(outputs)
```

```python
# WITHOUT UNSLOTH - NOT RECOMMENDED
from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer
model = AutoPeftModelForCausalLM.from_pretrained(
    "pavankumarbalijepalli/telLM-gemma2-9b",
    load_in_4bit = load_in_4bit,
)
tokenizer = AutoTokenizer.from_pretrained("pavankumarbalijepalli/telLM-gemma2-9b")

prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{}

### Input:
{}

### Response:
{}"""

inputs = tokenizer(
    [
        alpaca_prompt.format(
            "Naa peru enti?",  # instruction
            "Pavan Kumar ninnu aduguthunnadu",  # input
            "",  # output - leave this blank for generation!
        )
    ],
    return_tensors="pt",
).to("cuda")

outputs = model.generate(**inputs, max_new_tokens=64, use_cache=True)
tokenizer.batch_decode(outputs)
```

### Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->

__Generating Unintended Code:__

While the model can converse in telugu language, it may not be robust enough to handle complex logic or edge cases. Using it to generate critical production code could lead to errors or unexpected behavior in databases.

__Security Risks:__

`will_add_soon`.

__Beyond its Training Scope:__

The model is trained on telugu Language. Using it for a different languages could lead to inaccurate or nonsensical responses.

## Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->

__Bias and Fairness:__

The model's training data may contain biases that are reflected in the responses. This could lead to unfair or discriminatory outcomes, especially if the data is not carefully curated.

__Interpretability and Explainability:__

Language models are often "black boxes" where it's difficult to understand how they learn a specific language, in this case Telugu. This lack of interpretability makes it challenging to debug errors or ensure the generated queries are safe and efficient.

__Replacing Human Expertise:__

`will_add_soon`

### Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model.

## Training Details

### Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

```
@misc {divi_eswar_chowdary_2025,
author = { {Divi Eswar Chowdary} },
title = { Telugu_InstructData (Revision b734b7d) },
year = 2025,
url = { https://huggingface.co/datasets/eswardivi/Telugu_InstructData },
doi = { 10.57967/hf/4666 },
publisher = { Hugging Face }
}
```

## Evaluation

<!-- This section describes the evaluation protocols and provides the results. -->

### Testing Data, Factors & Metrics

#### Testing Data

<!-- This should link to a Dataset Card if possible. -->

Used eswardivi/Telugu_InstructData and split the data into training and testing datasets. The holdout dataset is used for testing the model. 

```
############# WILL UPDATE SOON #############
#### Factors

<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
The complexity of the questions are calculated using the number of tables per question, number of joins, group by, and sub queries per answer. This complexity is used to prepare the test data by stratifying the split around the complexity.

#### Metrics

<!-- These are the evaluation metrics being used, ideally with a description of why. -->
* __Execution Success:__ This metric is used to find out if the generated query is executable without arising any errors. For this, a sqllite3 connection is made to the memory, and using context the dummy tables are created. Then the predicted SQL is executed. This checks out if the generated query is in proper syntax, and if the model is hallucinating any new columns.
* __Inference Time:__ This metric is used to find out which model is providing results in less amount of time. This combined with the execution success, gives the efficiency of the model.
- 
### Results

* __Execution Success:__ Finetuned Phi-2 has 29% more success rate than the SQLCoder-7b-2
* __Inference Time:__ Finetuned Phi-2 has 41% increased inference speed than SQLCoder-7b-2  

#### Summary
* __Reduced Inference Time and Memory Footprint:__ The fine-tuned Phi-2 model
demonstrated a reduction in inference time and memory usage compared to the DeFog
SQLCoder. This is attributed to Phi-2's smaller size and the efficiency of quantization
techniques employed during fine-tuning. This finding implies that NL2SQL models can
be deployed on lower-powered devices like laptops or even mobile phones, potentially
democratizing access to this technology for a wider range of users.

* __Competitive Performance on Easy and Medium Queries:__ The fine-tuned Phi-2
achieved comparable performance to the DeFog SQLCoder in terms of accuracy on easy,
medium, and hard difficulty queries. This indicates that Phi-2, despite its smaller size,
can effectively handle a significant portion of real-world NL2SQL tasks, especially for
simpler queries.

* __Challenges with Complex Queries:__ While Phi-2 performed well on easier queries, it
encountered challenges with complex queries, exhibiting a drop in execution success
compared to the DeFog SQLCoder. This highlights the trade-off between model size and
complexity, suggesting that larger models might still be necessary for tackling highly
intricate tasks.

* __Potential for Further Improvement:__ The fine-tuning process employed in this study
can be further optimized by exploring different hyperparameter configurations and
potentially investigating alternative fine-tuning techniques like adapter-based methods.
This optimization has the potential to improve the model's performance on complex
queries while maintaining its efficiency.

```

## Environmental Impact

<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->

Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).

- **Hardware Type:** RTX 4070 PCIE 12GB X 1
- **Hours used:** 14 Hours
- **Cloud Provider:** Private Local Machine
- **Compute Region:** Asia-East-1
- **Carbon Emitted:** 1.81 kg eq. CO2


## Citation

<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**
```
@misc {pavan_kumar_balijepalli_2025,
	author       = { {Pavan Kumar Balijepalli} },
	title        = { telLM-gemma2-9b-16bit (Revision 7befd21) },
	year         = 2025,
	url          = { https://huggingface.co/pavankumarbalijepalli/telLM-gemma2-9b-16bit },
	doi          = { 10.57967/hf/4674 },
	publisher    = { Hugging Face }
}
```