pynb-73m-base

A 73M parameter language model trained for code generation with smolagents. Built on the Qwen2 architecture.

Model Details

Property	Value
Parameters	73.6M
Architecture	Qwen2ForCausalLM
Hidden size	384
Layers	12
Attention heads	6 (2 KV heads, GQA 3:1)
Intermediate size	768
Context length	2048
Vocab size	151,671

Training

Trained for 15,500 steps (~12 hours) on a single NVIDIA RTX 5070 Ti.

Metric	Start	End
Train Loss	12.0	2.4
Val Loss	6.5	2.6

Quick Start with smolagents

See inference_smolagent.py for full agent setup with LocalPythonExecutor and tools.

from inference_smolagent import create_agent, CalculatorTool, FibonacciTool

agent = create_agent(
    model_id="AutomatedScientist/pynb-73m-base",
    tools=[CalculatorTool(), FibonacciTool()],
    max_steps=5,
)

result = agent.run("Calculate 15 * 7 + 23")
print(result)

Or with HuggingFace API model:

from smolagents import CodeAgent, HfApiModel

model = HfApiModel(model_id="AutomatedScientist/pynb-73m-base")
agent = CodeAgent(tools=[], model=model)

result = agent.run("Calculate the sum of numbers from 1 to 100")
print(result)

Local Inference

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "AutomatedScientist/pynb-73m-base"  # or "checkpoint" for local
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

prompt = "Write a function to calculate fibonacci numbers"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, do_sample=True, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=False))

Inference Script

See inference.py for a wrapper class:

from inference import CodeModel

model = CodeModel("AutomatedScientist/pynb-73m-base")
result = model.generate("Write a function to sort a list")
print(result)

Installation

pip install torch transformers smolagents

Limitations

• Small model (73M params) - limited reasoning capacity compared to larger models
• Context window limited to 2,048 tokens
• Best used with short prompts due to context constraints

License

Apache 2.0