🤖 Safar-1.0

Safar-1.0 is a fine-tuned GPT-2 Medium (345M parameters) model trained on general English conversations and Computer Science educational content.

---

🧠 Model Details

• Base Model: GPT-2 Medium (345M parameters)
• Architecture: GPT2LMHeadModel
• Framework: PyTorch + Transformers
• Training Platform: Google Colab (T4 GPU)
• License: MIT (inherits GPT-2 license)

---

---

📊 Evaluation Results

Evaluated using EleutherAI lm-eval-harness (0-shot):

• ARC-Easy: 50.6%
• HellaSwag: 33.1%
• Winogrande: 53.2%

Performance is comparable to GPT-2 Medium baseline with minor improvements in reasoning consistency.

---

📚 Training Data

The model was fine-tuned on a mixture of open datasets including:

• General conversational text
• Computer Science educational material
• Programming-related discussions
• Academic-style explanations
• Public domain open datasets

---

⚙️ Training Details

• GPU: NVIDIA T4
• Training Time: ~25–45 minutes
• Optimizer: AdamW
• Fine-tuned using HuggingFace Trainer
• Final Training Loss: (add your value here)

---

💬 Intended Use

Safar-1.0 is designed for:

• General chat
• CS topic explanations
• Educational Q&A
• English text generation

---

⚠️ Limitations

• May hallucinate factual information
• Not aligned with RLHF
• Not suitable for critical production usage
• Performance lower than modern large LLMs

---

🚀 Example Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "ujjwaljha/Safar-1.0"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "Explain binary search in simple terms."

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=100)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

📊 Model Type

💬 Chat Model (Instruction-style fine-tuned GPT-2)

---

👤 Author

Built and trained by Ujjwal Jha.