README.md

---
library_name: transformers
tags:
- code-generation
- concept-embedding
- jepa
- pytorch
- gguf
license: apache-2.0
---

# Concept-First Code Generation

**Inspired by VL-JEPA**: Predict concept embeddings first, then generate code conditioned on them.

## The Idea

Traditional autoregressive models predict tokens one at a time, which can lead to losing coherence or hallucinating APIs. The **Concept-First** approach solves this by:

1. **Concept Encoder**: Encoding code snippets into semantic embeddings.
2. **Concept Predictor**: Predicting what the code embedding should look like given a query.
3. **Concept-Conditioned Generation**: Retrieving similar concepts to guide the LLM.

```mermaid
graph LR
    A[Query] --> B(Concept Predictor)
    B --> C{Concept Space}
    C --> D[Retrieve Similar Code]
    D --> E[Conditioned Generation]
```

## Models Used (January 2026)

| Component | Model | Description |
|-----------|-------|-------------|
| **Concept Encoder** | `Salesforce/SFR-Embedding-Code-2B_R` | SOTA code embeddings (CoIR: 67.4) |
| **Text Encoder** | `Alibaba-NLP/gte-Qwen2-1.5B-instruct` | State-of-the-art text embedding |
| **Concept Predictor** | Custom MLP | Maps text queries to code concept space |
| **Code LLM** | `Qwen/Qwen2.5-Coder-32B-Instruct` | High-performance code generation |

## Files in this Repo

- `concept_predictor.pt`: PyTorch weights for the concept predictor MLP.
- `concept_predictor.gguf`: GGUF format for edge deployment (llama.cpp/LM Studio).
- `concept_bank.pt`: Pre-computed embeddings for the concept retrieval bank.

## Usage

```python
# Load the concept predictor
import torch
checkpoint = torch.load("concept_predictor.pt")
# ... (See Colab notebook for full implementation)
```

## Datasets

Constructed from high-quality subsets of:
- **MBPP**
- **Evol-Instruct-Code**
- **Magicoder-OSS-Instruct**

## Credits

Created by **Core Subagent** (Colab Composer) for **Riley Seaburg**.