README.md

---
license: mit
tags:
  - gpt
  - hypernetwork
  - lora
  - adaptive-attention
  - diffusion
  - language-model
  - from-scratch
language:
  - en
---

# Adaptive GPTs

Five small GPT variants trained from scratch for 12k steps on a ~300M token mixed corpus (TinyStories, Cosmopedia WikiHow, WikiText-2, Python code). The core experiment: can a hypernetwork that reads the residual stream generate dynamic LoRA-style updates to the value heads at inference time?

**Blog post:** [Adaptive Attention at Inference Time: Does It Actually Work?](https://teendifferent.substack.com/p/adaptive-attention-at-inference-time)

**Code:** [REDDITARUN/a-gpt](https://github.com/REDDITARUN/a-gpt)

## Checkpoints

| File | Model | Params | Description |
|---|---|---|---|
| `base_best.pt` | Base GPT | 28.9M | Vanilla causal GPT. 4 layers, 4 heads, 256 dim. |
| `matched_best.pt` | Matched GPT | 30.5M | 6 layers, 4 heads, 256 dim. Parameter-matched to adaptive. |
| `adaptive_best.pt` | Adaptive GPT | 30.5M | Base + per-head TinyHeadTransformer hypernetwork on V. |
| `diffusion_best.pt` | Diffusion GPT | 28.9M | Bidirectional denoising (discrete diffusion). |
| `adaptive_diffusion_best.pt` | Adaptive Diffusion GPT | 30.5M | Diffusion + per-head hypernetwork on V. |

## Training

All models share: AdamW, lr 3e-4 with cosine decay, batch size 8, 4 gradient accumulation steps, 256 context window, GPT-2 tokenizer, RoPE.

Data mixture: Cosmopedia WikiHow (40%), Python code (30%), WikiText-2 (20%), TinyStories (10%).

## Loading

```python
import torch
from models.base_gpt import GPT_Base, GPT_Base_Config
from models.a_gpt import GPT_Custom, GPT_Custom_Config, HyperConfig

# Example: load adaptive checkpoint
config = GPT_Custom_Config()
hyper_config = HyperConfig()
model = GPT_Custom(config, hyper_config=hyper_config)

ckpt = torch.load("adaptive_best.pt", map_location="cpu")
model.load_state_dict(ckpt["model"])