Airavat-ai/erawan-1

Erawan 1

Erawan 1 is a 1.7B-parameter language model fine-tuned for reasoning and long-horizon agentic tasks, developed by Airavat AI Micro Lab. Built on Qwen3-1.7B, fine-tuned with QLoRA and merged into a standalone full-precision checkpoint.

Model Details

• Developed by: Airavat AI Micro Lab
• Base model: Qwen3-1.7B (1,755,440,128 parameters)
• Architecture: Qwen3 decoder-only transformer, 28 layers, grouped-query attention
• Fine-tuning method: QLoRA (4-bit base, LoRA r=32) via Unsloth, merged into fp16 weights
• Training steps: 260 cumulative fine-tune steps
• Context length: 4096 tokens
• License: Apache 2.0

Benchmark Results (Erawan 1 vs base Qwen3-1.7B)

Task	Benchmark	Base	Erawan 1
Reasoning / Maths	GSM8K (n=10)	1/10	3/10
Coding / Logic	HumanEval (n=6)	4/6	4/6
Tool-use	Glaive function-calling (n=6)	6/6	6/6

HumanEval vs Public Small Models

Model	Params	HumanEval
Llama-3.2-1B-Instruct	1B	33.5%
SmolLM2-1.7B-Instruct	1.7B	28.1%
Qwen2.5-1.5B-Instruct	1.5B	30.5%
Qwen3-1.7B (base)	1.7B	67%
Erawan 1	1.7B	67%

Training Data

• unsloth/OpenMathReasoning-mini (~600 rows)
• NousResearch/hermes-function-calling-v1 (~900 rows)
• mlabonne/FineTome-100k (~200 rows)
• Custom identity examples

Intended Use

Agentic tool-calling workflows, step-by-step reasoning, and research on small efficient reasoning models. Not intended for medical, legal, or financial decision-making, or any use requiring guaranteed factual accuracy.

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained('airavat-ai/erawan-1')
tokenizer = AutoTokenizer.from_pretrained('airavat-ai/erawan-1')
messages = [{'role': 'user', 'content': 'Who are you?'}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors='pt')
outputs = model.generate(inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Credits

Base model: Qwen3-1.7B by the Qwen Team, Alibaba Cloud. Fine-tuning tooling: Unsloth. License: Apache 2.0.