11-47/GPT2.5.2-NSFW-Codex-0.4B

GPT2.5.2-NSFW-Codex-0.4B

📌 Model Overview

Model Name: WithinUsAI/GPT2.5.2-NSFW-Codex-0.4B Organization: Within Us AI Base Model: openai-community/gpt2-medium (~0.4B parameters)  Model Type: Instruction-Tuned Code + General Text LLM Parameter Size: ~0.4B Primary Focus: Lightweight coding + uncensored responses + dataset-driven evolution

This model represents a heavily evolved GPT-2 Medium, upgraded through dataset-driven training and reasoning/coding distillation.

🧬 Name Meaning

“GPT2.5.2” = Evolution lineage

• —2 → GPT-2 Medium base
• —.5.2 → GPT 5.2 Codex-style

👉 In short: A GPT-2 model pushed forward toward modern coding + reasoning behavior using curated datasets, not architecture scaling.

⸻

🧬 Architecture & Lineage

Base Foundation

• —Transformer architecture from GPT-2 Medium (~345M–400M parameters class) 
• —Dense, autoregressive language model

Evolution Process

This model was scaled up,evolved through data and training strategy:

• —Instruction tuning
• —Coding dataset exposure (Codex-style tasks)
• —Reasoning trace influence
• —Behavioral refinement toward modern LLM outputs

⸻

🧠 Core Design Philosophy

Don’t scale the model… evolve it.

Instead of increasing parameters, this model:

• —Improves behavior through data quality
• —Mimics GPT 5.2 Codex reasoning styles
• —Pushes GPT-2 into modern task domains

Think of it like a classic engine rebuilt with modern parts 🔧

⸻

⚙️ Key Capabilities

💻 Coding (Codex-Inspired)

• —Basic code generation (Python, JS, etc.)
• —Simple debugging assistance
• —Structured function outputs

🧠 Reasoning (Lightweight)

• —Step-by-step responses (limited depth)
• —Instruction-following improvements over base GPT-2

🔓 Uncensored Behavior

• —Reduced refusal tendencies
• —More permissive outputs compared to aligned models

⸻

📦 Model Characteristics

Attribute Value Parameters ~0.4B Architecture GPT-2 (dense transformer) Strength Efficiency Weakness Limited deep reasoning

⸻

🚀 Intended Use

✅ Ideal Use Cases

• —Ultra-lightweight local LLMs
• —Experimental coding assistants
• —Dataset-driven model research
• —Uncensored response exploration
• —Edge/low-resource environments

⚠️ Limitations

• —Significantly weaker than modern 7B+ models
• —Limited context and reasoning depth

⸻

🛠️ Example Usage (Transformers)

from transformers import AutoTokenizer, AutoModelForCausalLM modelname = "WithinUsAI/GPT2.5.2-NSFW-Codex-0.4B" tokenizer = AutoTokenizer.frompretrained(modelname) model = AutoModelForCausalLM.frompretrained(modelname) prompt = "Write a Python function to reverse a string." inputs = tokenizer(prompt, returntensors="pt") outputs = model.generate(**inputs, maxnewtokens=100) print(tokenizer.decode(outputs[0]))

⸻

🧪 Training Methodology

Within Us AI approach:

• —Dataset-driven evolution (primary driver)
• —Instruction + coding task fine-tuning
• —Reasoning-style output shaping

Data Sources

• —Proprietary datasets created by Within Us AI
• —Third-party datasets used without ownership claims
• —Likely includes:
• —Coding tasks
• —Instruction datasets
• —Reasoning traces

⸻

📊 Expected Performance Profile

Capability Strength Efficiency Very High Coding (basic) High Reasoning High Creativity Moderate Safety filtering Minimal

⸻

📜 License

License Type: Based on GPT-2 (OpenAI open model lineage)**

Attribution Notes:

• —Base model: OpenAI GPT-2 Medium
• —Training / evolution: Within Us AI
• —Third-party datasets used without ownership claims
• —Credit belongs to original dataset and model creators

⸻

🙏 Acknowledgements

• —OpenAI (GPT-2 architecture)
• —Hugging Face (model hosting ecosystem)
• —Open-source dataset contributors
• —Coding dataset communities

⸻

🔗 Links

• —Model: https://huggingface.co/WithinUsAI/GPT2.5.2-NSFW-Codex-0.4B
• —Organization: https://huggingface.co/WithinUsAI

⸻

🧩 Closing Note

This model is like a time traveler from 2019 carrying tools from 2026 ⏳⚡

Same small brain… but trained to think like GPT 5.2 Codex.