---
language: en
license: other
base_model: Qwen/Qwen2.5-3B-Instruct
tags:
- vyasa
- architectural-reasoning
- structured-generation
- code-generation
- lora
- peft
datasets:
- prashantpandey/vyasa-architectural-pairs
metrics:
- valid-json-rate
- witness-clean-rate
---

# Vyasa-Architect-3B

A 3B-parameter language model fine-tuned for architectural reasoning — converting natural language app descriptions into complete, structured Architectural Graph JSON.

Named after Vyasa (व्यास), the sage who conceived the entire Mahabharata completely in his mind before manifesting a single word.

## Model Description

- **Base Model:** Qwen-2.5-3B-Instruct (4-bit quantized)
- **Fine-tuning Method:** QLoRA (LoRA rank 8, 8 layers, 3.33M trainable params)
- **Training Data:** 167 (intent → Architectural Graph) pairs generated by DeepSeek-V3
- **Training Hardware:** Apple M1 Pro, 32GB unified memory
- **Training Duration:** ~1 hour, 300 iterations
- **Training Loss:** 0.063 (train), 0.124 (validation)

## Intended Use

This model is the Stage 1 (Conception) component of the Vyasa Manifestation Engine. It receives a natural language app description and produces a complete Architectural Graph JSON. This graph is then validated by a deterministic Witness (12 structural checks) and compiled to working code files by the Manifest Engine (deterministic Next.js/Prisma/TypeScript compiler).

The model does NOT generate code directly. It generates architectural specifications. Code is produced deterministically by the compiler.

## Performance

| Metric (n=50, preregistered protocol) | Base Qwen-2.5-3B | Vyasa-Architect-3B (this adapter) |
|--------|------------------|--------------------|
| Parseable JSON | 62% | 40% |
| Schema-valid graph | 8% | 38% |
| Witness-clean | 6% [2.1–16.2] | 14% [7.0–26.2] |

Measured 2026-07-03 (greedy decoding, Wilson 95% CIs). Fine-tuning teaches
schema conformance (8% → 38%) and structural discipline (6% → 14%), not JSON
syntax — the untrained base already parses 62% of the time. A successor trained
on concise pairs (arch-v2 recipe) measures 34% Witness-clean, 52% after a
deterministic Witness-guided repair pass. GRPO+Witness training is gated on a
preregistered pass@8 headroom probe.

## Usage

```python
from mlx_lm import load, generate

model, tokenizer = load(
    "mlx-community/Qwen2.5-3B-Instruct-4bit",
    adapter_path="prashantpandey/vyasa-architect-3b"
)

prompt = "Conceive the complete architecture for: A todo app with tasks and categories"
response = generate(model, tokenizer, prompt=prompt, max_tokens=3000)
```

## Architecture

Part of the Vyasa Manifestation Engine — a three-stage compiler architecture:
1. **Conception** (this model): Intent → Architectural Graph
2. **Witness** (deterministic): Validates structural integrity
3. **Manifest** (deterministic): Compiles graph to code files

## Limitations

- 40% parseable / 14% Witness-clean at n=50 (see Performance) — verbosity-truncation at long outputs is the dominant failure
- Output is verbose (8K-11K chars per graph)
- Slow on CPU (~200 chars/sec); optimized for GPU inference
- Trained for Next.js/Prisma stack output; graph format is stack-agnostic

## Citation

If you use this model, please cite:
```
@misc{vyasa-architect-3b,
  author = {Prashant Pandey},
  title = {Vyasa-Architect-3B: Architectural Reasoning via Structured Intermediate Representation},
  year = {2026},
  url = {https://github.com/prashantpandey-creator/puranic-architecture-paper}
}
```

## License

This adapter is proprietary. All rights reserved. The base model (Qwen-2.5-3B) is Apache 2.0.