112 lines
3.0 KiB
Markdown
112 lines
3.0 KiB
Markdown
|
---
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
license: mit
|
|||
|
|
tags:
|
|||
|
|
- gpt2
|
|||
|
|
- causal-lm
|
|||
|
|
- text-generation
|
|||
|
|
- distillation
|
|||
|
|
- fine-tuned
|
|||
|
|
- withinusai
|
|||
|
|
- full-fine-tune
|
|||
|
|
- thinking
|
|||
|
|
- reasoning
|
|||
|
|
base_model: openai-community/gpt2
|
|||
|
|
datasets:
|
|||
|
|
- WithinUsAI/GPT_5.5_Distilled
|
|||
|
|
- WithinUsAI/GPT5.5_thinking_max_distill_god_seed_25K
|
|||
|
|
model_type: gpt2
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
library_name: transformers
|
|||
|
|
widget:
|
|||
|
|
- text: "The key to understanding intelligence is"
|
|||
|
|
- text: "Let me think through this step by step:"
|
|||
|
|
- text: "Deep within the layers of thought,"
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# GPT2.5.5-Awakened.Thinker-0.1B
|
|||
|
|
|
|||
|
|
> *Forged by WithIn Us AI — a fully fine-tuned GPT-2 awakened on distilled GPT-5.5 thinking patterns.*
|
|||
|
|
|
|||
|
|
## Model Overview
|
|||
|
|
|
|||
|
|
| Property | Value |
|
|||
|
|
|---|---|
|
|||
|
|
| **Architecture** | GPT-2 (`openai-community/gpt2`) |
|
|||
|
|
| **Parameters** | ~124M (0.1B class) |
|
|||
|
|
| **Training Type** | Full fine-tune — ALL weights updated, zero adapters |
|
|||
|
|
| **Context Window** | 1024 tokens |
|
|||
|
|
| **Best Eval Loss** | `0.4365` |
|
|||
|
|
| **Best Perplexity** | `1.55` |
|
|||
|
|
| **Creator** | [GODsStrongestSoldier](https://huggingface.co/GODsStrongestSoldier) / WithIn Us AI |
|
|||
|
|
| **Date Trained** | 2026-05-23 |
|
|||
|
|
| **Hardware** | 2× NVIDIA Tesla T4 (Kaggle) |
|
|||
|
|
| **Precision** | FP16 mixed precision |
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Training Methodology
|
|||
|
|
|
|||
|
|
Full fine-tuning — every single parameter in GPT-2 was updated.
|
|||
|
|
No LoRA, no QLoRA, no adapters of any kind.
|
|||
|
|
|
|||
|
|
### Datasets
|
|||
|
|
|
|||
|
|
| Dataset | Description |
|
|||
|
|
|---|---|
|
|||
|
|
| [`WithinUsAI/GPT_5.5_Distilled`](https://huggingface.co/datasets/WithinUsAI/GPT_5.5_Distilled) | Instruction + completion pairs distilled from GPT-5.5 |
|
|||
|
|
| [`WithinUsAI/GPT5.5_thinking_max_distill_god_seed_25K`](https://huggingface.co/datasets/WithinUsAI/GPT5.5_thinking_max_distill_god_seed_25K) | 25K chain-of-thought reasoning traces distilled from GPT-5.5 |
|
|||
|
|
|
|||
|
|
97 / 3 train / eval split.
|
|||
|
|
|
|||
|
|
### Hyperparameters
|
|||
|
|
|
|||
|
|
| Parameter | Value |
|
|||
|
|
|---|---|
|
|||
|
|
| Peak Learning Rate | `3e-5` |
|
|||
|
|
| LR Schedule | Cosine with 6% warmup |
|
|||
|
|
| Effective Batch Size | `64` (4 × 2 GPUs × 8 grad accum) |
|
|||
|
|
| Epochs | `5` |
|
|||
|
|
| Weight Decay | `0.1` |
|
|||
|
|
| Max Sequence Length | `1024` |
|
|||
|
|
| Precision | FP16 |
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Quick Start
|
|||
|
|
|
|||
|
|
```python
|
|||
|
|
from transformers import GPT2LMHeadModel, GPT2TokenizerFast
|
|||
|
|
import torch
|
|||
|
|
|
|||
|
|
model_id = "GODsStrongestSoldier/GPT2.5.5-Awakened.Thinker-0.1B"
|
|||
|
|
tokenizer = GPT2TokenizerFast.from_pretrained(model_id)
|
|||
|
|
model = GPT2LMHeadModel.from_pretrained(model_id, torch_dtype=torch.float16)
|
|||
|
|
model.eval()
|
|||
|
|
|
|||
|
|
prompt = "Let me think through this carefully, step by step:"
|
|||
|
|
inputs = tokenizer(prompt, return_tensors="pt")
|
|||
|
|
|
|||
|
|
with torch.no_grad():
|
|||
|
|
output = model.generate(
|
|||
|
|
**inputs,
|
|||
|
|
max_new_tokens = 200,
|
|||
|
|
do_sample = True,
|
|||
|
|
temperature = 0.7,
|
|||
|
|
top_p = 0.9,
|
|||
|
|
repetition_penalty = 1.15,
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
print(tokenizer.decode(output[0], skip_special_tokens=True))
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## About WithIn Us AI
|
|||
|
|
|
|||
|
|
- HuggingFace org : [WithinUsAI](https://huggingface.co/WithinUsAI)
|
|||
|
|
- Creator : [GODsStrongestSoldier](https://huggingface.co/GODsStrongestSoldier)
|
|||
|
|
|
|||
|
|
*"Strength through understanding. Awakened from within."*
|