README.md

---
license: mit
tags:
- effect
- typescript
- reinforcement-learning
- grpo
- fine-tuning
- qwen
- unsloth
- lora
---

# Qwen3-4B — Effect TypeScript Code Generation

Fine-tuned Qwen3-4B model specialized in generating high-quality **Effect-style TypeScript code** using Reinforcement Learning (GRPO).

## Model Details

- **Developed by**: Kodep
- **Model type**: Qwen3-4B with LoRA adapter (rank 64)
- **Language**: TypeScript (Effect framework)
- **License**: MIT
- **Base model**: [unsloth/Qwen3-4B](https://huggingface.co/unsloth/Qwen3-4B)
- **Repository**: [github.com/belarusian/training](https://github.com/belarusian/training)

## What is this model?

This model generates Effect-style TypeScript code — the popular effect system for functional programming in TypeScript. It's been fine-tuned using **GRPO** (Group Relative Policy Optimization), a reinforcement learning algorithm that improves code quality through reward-based training.

The model handles:
- Effect imports and core patterns (`Effect.succeed`, `Effect.flatMap`, etc.)
- Effect Schema definitions (`Schema`, `decodeSync`, etc.)
- Effect service patterns
- Proper TypeScript exports and types

## How to Use

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("Kodep/qwen3-4b-effect-codegen")
model = AutoModelForCausalLM.from_pretrained(
    "Kodep/qwen3-4b-effect-codegen",
    torch_dtype=torch.float16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are an expert TypeScript developer specializing in the Effect framework."},
    {"role": "user", "content": "Generate an Effect service pattern for a user repository"},
]

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, max_new_tokens=1024, temperature=0.7)
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

### With Unsloth (faster inference)

```python
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="Kodep/qwen3-4b-effect-codegen",
    max_seq_length=4096,
    load_in_4bit=True,
)

# Inference
messages = [
    {"role": "system", "content": "You are an expert TypeScript developer specializing in the Effect framework."},
    {"role": "user", "content": "Generate an Effect Effect pattern for a user repository"},
]
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, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

## Training Details

### Training Data
- **428** TypeScript code samples extracted from:
  - `effect-smol` — 185 samples
  - `effect` — 208 samples
  - `opencode` — 28 samples
  - `effect-examples` — 7 samples
- Sources: Real Effect.js library code, OpenCode LLM integrations, and Effect examples

### Training Procedure
1. **Data extraction**: TypeScript files scraped from Effect repositories, filtered for Effect-specific imports
2. **SFT pre-training** (optional): 2 epochs, learning rate 2e-4 — teaches code format
3. **GRPO training**: 1 epoch, learning rate 2e-6, batch size 1, gradient accumulation 4x
4. **Reward functions**:
   - **+1.0** Code has `<CODE>` tags
   - **+0.5** Has Effect imports
   - **+0.3** Uses Schema
   - **+0.2** Has exports
   - **-0.5** Response too short (<100 chars)

### Hyperparameters
| Parameter | Value |
|-----------|-------|
| Base model | Qwen3-4B |
| LoRA rank | 64 |
| Max sequence | 4096 |
| SFT lr | 2e-4 |
| GRPO lr | 2e-6 |
| SFT epochs | 2 |
| GRPO epochs | 1 |
| Optimizer | adamw_8bit |
| Gradient accum. | 4 |

### Hardware
- **GPU**: NVIDIA GeForce RTX 4090 (24GB VRAM)
- **CUDA**: 13.0
- **PyTorch**: 2.10.0+cu130
- **Unsloth**: 2026.5.8
- **vLLM**: Used for faster inference during GRPO

## Risks and Limitations

- Fine-tuned on a small dataset (428 samples) — may not cover all Effect patterns
- May generate syntactically valid but logically incorrect code
- Not suitable for production use without evaluation
- Training focused on code format and import patterns, not correctness verification

## Citation

```bibtex
@misc{qwen3-4b-effect-codegen,
  author = {Kodep},
  title = {Qwen3-4B Effect TypeScript Code Generation},
  year = {2026},
  url = {https://huggingface.co/Kodep/qwen3-4b-effect-codegen}
}
```

## Related

- [Effect TypeScript Library](https://effect.website/)
- [GRPO Paper](https://arxiv.org/abs/2402.03300)
- [Unsloth](https://github.com/unslothai/unsloth)
- [Training Repository](https://github.com/belarusian/training)