README.md

---
base_model:
- deepseek-ai/deepseek-coder-1.3b-base
finetuned version: aiswaryards/deepseek1.3B-coder-dora-finetuned

tags:
- '`DoRA`'
- '`question-generation`'
- '`data-science`'
- '`knowledge-transfer`'
- '`rag`'
- '`llm`'
- '`agents`'
- '`deepseek`'

model_type: causal-lm
---
## Quick Overview:

The deployed model `aiswaryards/deepseek1.3B-coder-dora-finetuned` is a fine-tuned version of [DeepSeek-Coder 1.3B](https://huggingface.co/deepseek-ai/deepseek-coder-1.3b-base), adapted using the **DoRA (Decoupled Low-Rank Adaptation)** method.

It was trained as part of a **graduate-level academic project** focused on simulating expert question generation from domain-specific Knowledge Transfer (KT) documents.

To fine-tune a compact and capable LLM to generate **high-quality, context-specific technical questions** based on internal handover documents in a Retrieval-Augmented Generation (RAG) pipeline.


  
## Use Case
- Designed to simulate domain expert handover, where the model generates precise and context-aware questions to extract undocumented insights from KT documents.
- Create prompts that makes the 

## Model Information:
Base: deepseek-ai/deepseek-coder-1.3b-base

Fine-tuning: (DoRA) Dynamically Optimized Rank Adaptation

Format: Instruction → Input (context) → Response (question)

## Fine-Tuning Details

- **Base Model**: `deepseek-ai/deepseek-coder-1.3b-base`

- **Adapter Method**: DoRA using `AdaLoraConfig` from PEFT
  - `r`: 8
  - `lora_alpha`: 16
  - `lora_dropout`: 0.05
  - `bias`: none
  - `target_modules`: ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]
  - use_dora=True

- **Quantization**: bfloat16

- **Data Source**: Webscrapped dataset (~100+ high-quality Q&A pairs from Knowledge Transfer docs)

- **Training**:
  - Epochs: 5
  - Batch Size: 1 (gradient accumulation: 8)
  - Optimized using Hugging Face `Trainer`
  - Platform: Google Colab (A100)
  - Quantization: 16-bit (bfloat16)
  - Hardware: Google Colab - (A100 GPU) - cuda
  - Frameworks: HuggingFace Transformers, PEFT, Datasets, Trainer
  - Tokenization length: 1024
  - Trained on :  high-quality Q&A pairs collected from various branches of data science domain KT handovers.


## Training Performance:

 Step |	Train Loss  |	Validation Loss
 =======================================
 400  |	   1.780	|        1.795      
 =======================================
- This exhibits a strong convergence for a fine-tuned instruction. 
- The validation loss curve flattens under 2.0, which is a great sign for generative quality.


## Use Cases

- Domain Expert Simulation (Agentic RAG)
- Knowledge Transfer Automation
- Multiple roles - currently experimented on datascience domain (for eg. BI / Data Engineering role transitions)
- Question synthesis for downstream QA chains

## Challenges:

- It requires instruction-style prompting. 
- Might hallucinate if context is vague or unrelated  
- It is best suited within structured RAG pipelines

## License

---
license: deepseek
---

### DeepSeek-Coder 1.3B - DoRA Fine-tuned Version

This model is a fine-tuned version of `deepseek-ai/deepseek-coder-1.3b-base` using the DoRA method on academic Q&A data from Knowledge Transfer documents.

...

This model inherits the original license from [DeepSeek-AI](https://huggingface.co/deepseek-ai/deepseek-coder-1.3b-base), which can be found [here](https://huggingface.co/deepseek-ai/deepseek-coder-1.3b-base/blob/main/LICENSE).

For academic use only. Please refer to the original license terms for reuse, distribution, or commercial applications.