---
library_name: transformers
tags:
  - code
  - coding-assistant
  - qwen2
  - lora
  - fine-tuned
  - full-stack
  - reasoning
license: apache-2.0
language:
  - en
base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct
pipeline_tag: text-generation
---

# 🇮🇳 IndraCoder — AI Coding Assistant

A fine-tuned coding LLM built on **Qwen2.5-Coder-1.5B-Instruct**, trained on 4 curated datasets for code generation, debugging, algorithmic reasoning, and agentic tool use.

## ✨ Highlights

- 🧠 **Chain-of-thought reasoning** — Uses `<think>` blocks to reason before coding
- 🔧 **Full-stack development** — Python, JavaScript, TypeScript, React, FastAPI, and more
- 🛠️ **Tool/function calling** — Trained on agentic tool-use patterns
- 📦 **Lightweight** — 1.5B parameters, runs on consumer GPUs

## Quick Start

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("RockySinghRajput/Indracoder", torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("RockySinghRajput/Indracoder")

messages = [
    {"role": "system", "content": "You are IndraCoder, an expert AI coding assistant."},
    {"role": "user", "content": "Write a Python function to find the longest palindromic substring."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

output = model.generate(inputs.input_ids, max_new_tokens=512, temperature=0.7, top_p=0.9)
print(tokenizer.decode(output[0][len(inputs.input_ids[0]):], skip_special_tokens=True))
```

## Model Details

| Property | Value |
|----------|-------|
| **Base Model** | [Qwen/Qwen2.5-Coder-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct) |
| **Parameters** | 1.5B |
| **Type** | Causal Language Model (merged LoRA fine-tune) |
| **Language** | English |
| **License** | Apache 2.0 |
| **Developed by** | [RockySinghRajput](https://huggingface.co/RockySinghRajput) |

## Training Details

### Training Data

Fine-tuned on **4 curated datasets** (~8,000 samples):

| Dataset | Purpose | Samples |
|---------|---------|---------|
| [glaive-code-assistant-v3](https://huggingface.co/datasets/glaiveai/glaive-code-assistant-v3) | General code generation & debugging | ~2,000 |
| [evol-codealpaca-v1](https://huggingface.co/datasets/theblackcat102/evol-codealpaca-v1) | Hard algorithmic problems | ~2,000 |
| [CodeFeedback-Filtered](https://huggingface.co/datasets/m-a-p/CodeFeedback-Filtered-Instruction) | Code reasoning & explanations | ~2,000 |
| [glaive-function-calling-v2](https://huggingface.co/datasets/glaiveai/glaive-function-calling-v2) | Agentic tool/function calling | ~2,000 |

### Training Procedure

- **Method**: LoRA (Low-Rank Adaptation) → merged into base model
- **LoRA Config**: r=16, alpha=16, dropout=0.05
- **Target Modules**: q_proj, k_proj, v_proj, o_proj
- **Epochs**: 1
- **Batch Size**: 1 (gradient accumulation: 4, effective batch: 4)
- **Learning Rate**: 1e-4 (cosine schedule)
- **Optimizer**: paged_adamw_8bit
- **Sequence Length**: 512 tokens
- **Precision**: FP16 mixed precision
- **Quantization**: 4-bit NF4 (QLoRA) during training

### Compute Infrastructure

- **Hardware**: NVIDIA T4 GPU
- **Training Time**: ~1 hour

## Capabilities

### ✅ What IndraCoder Can Do

- **Write code** in Python, JavaScript, TypeScript, Java, C++, Go, Rust
- **Debug code** — find and fix bugs with explanations
- **Explain code** — break down complex code step by step
- **Algorithm design** — data structures, dynamic programming, graphs
- **Full-stack development** — React, FastAPI, Express, databases
- **Tool/function calling** — structured function calls for agentic workflows

### ⚠️ Limitations

- **1.5B model** — smaller than GPT-4, Claude, or larger open-source models
- **Not suitable** for complex multi-file refactoring or very long code generation
- **English only** — not trained on multilingual data
- **No image/file understanding** — text-only model
- **May hallucinate** — always review generated code before using in production

### ❌ Out-of-Scope Use

- Production code without human review
- Security-critical applications without expert validation
- Medical, legal, or financial advice
- Generating malicious code or exploits

## Evaluation

Tested on 4 qualitative benchmarks:

| Test | Task | Result |
|------|------|--------|
| Full-Stack | REST API with auth in FastAPI | ✅ Generates working code |
| Algorithm | Implement LRU Cache O(1) | ✅ Correct approach |
| Debug | Fix React infinite re-render | ✅ Identifies useEffect issue |
| Tool Use | Chain function calls for file analysis | ✅ Correct tool selection |

> **Note**: These are qualitative assessments, not standardized benchmarks.

## Citation

```bibtex
@misc{indracoder2025,
  title={IndraCoder: A Fine-tuned Coding LLM},
  author={RockySinghRajput},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/RockySinghRajput/Indracoder}
}
```

## Contact

- **HuggingFace**: [RockySinghRajput](https://huggingface.co/RockySinghRajput)