README.md

metadata

license: llama3.2
language:
  - en
base_model:
  - meta-llama/Llama-3.2-1B
pipeline_tag: text-generation

Model Card for InfiR-1B-Base

InfR aims to advance AI systems by improving reasoning, reducing adoption barriers, and addressing privacy concerns through smaller model sizes.

Model Details

Model Description

• Developed by: InfiX
• Language(s) (NLP): English
• Continual pretrained from model: [meta-llama/Llama-3.2-1B]

Model Sources

• Repository: [github]
• Paper [optional]: [Arxiv]

Uses

Bias, Risks, and Limitations

• Base model limitations: This is a base model that primarily performs text completion rather than instruction following. It may not understand complex questions or provide direct answers as effectively as instruction-tuned models. For better question-answering capabilities, consider using the instruction-tuned version (InfiR-1B-Instruct).
• Performance gaps remain vs. 70 B+ models on very hard reasoning (e.g., OlympiadBench).
• Safety & bias: inherits Llama-3.2 tokenizer & pre-training distribution; may reflect web biases.
• Knowledge cut-off: mid-2023.
• Evaluation has focused on English benchmarks; multilingual robustness not verified.

How to Get Started with the Model

Installation

First, install the required dependencies:

pip install torch transformers

For optimal performance, we recommend using PyTorch 2.0+ and CUDA 11.8+.

Basic Usage

Here's a simple example to get started with InfiR-1B-Base:

from transformers import AutoTokenizer, AutoModelForCausalLM

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("InfiX-ai/InfiR-1B-Base")
model = AutoModelForCausalLM.from_pretrained("InfiX-ai/InfiR-1B-Base")

# Example prompt
prompt = r"A new program had 60 downloads in the first month. The number of downloads in the second month was three times as many as the downloads in the first month, but then reduced by 30% in the third month. How many downloads did the program have total over the three months?"

# Tokenize and generate
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=2048)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Advanced Usage Examples

1. Mathematical Reasoning

# Mathematical problem solving
math_prompt = """Solve this step by step:

Problem: If a rectangle has a length of 8 units and a width of 6 units, what is its area and perimeter?

Solution:"""

inputs = tokenizer(math_prompt, return_tensors="pt")
outputs = model.generate(
    **inputs, 
    max_new_tokens=512,
    temperature=0.1,
    do_sample=True
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

2. Code Generation

# Code generation example
code_prompt = """Write a Python function to calculate the factorial of a number:

def factorial(n):
"""

inputs = tokenizer(code_prompt, return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.2,
    do_sample=True
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

3. Chain-of-Thought Reasoning

# Chain-of-thought reasoning
cot_prompt = """Let's approach this step by step:

Question: A train travels 120 km in 2 hours. What is its speed in km/h?

Let me think through this:"""

inputs = tokenizer(cot_prompt, return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_new_tokens=300,
    temperature=0.3,
    do_sample=True
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Training Data

Stage	Tokens	Composition
Pre-training	900 B	52 % code, 48 % high-quality web (math, science, encyclopedic)
Annealing	40 B	extra math & code + synthetic samples
SFT	~4 M	Infinity-Instruct, Orca-AgentInstruct-1M, NuminaMath, ScaleQuest (filtered)

Data cleaning: heuristic filters, MinHash de-duplication, 10-gram benchmark decontamination, reward-model rejection sampling.

Training Procedure

Hyper-parameter	Value
Precision	bf16 mixed
Optimizer	AdamW
LR (pre-train)	1.4 e-3, cosine → 0
LR (SFT)	2 e-5, cosine w/ 10 % warm-up
Batch size	2048 (pre-train), 128 (SFT)
Sequence len	4096
Epochs	1 (pre-train), 1 (anneal), 4 (SFT)
GPUs	64 × H800, 5760 GPU-hours total

Evaluation

Benchmarks & Results

Benchmark	InfiR-1B-Base	Llama-3.2-1B	Qwen-2.5-1.5B
MMLU	47.24	32.74	63.03
GSM8K	63.46	8.11	66.57
MATH	31.82	3.42	31.24
HumanEval	37.80	17.68	35.37
MBPP	53.40	49.0	58.37
MBPP(3-shot)	37.6	24.8	41.4

Technical Specifications

Model Architecture and Objective

• Base: Llama-3.2-1B (32 layers, 32 heads, RoPE, GQA, 2 k ctx → 4 k extended)

Citation

BibTeX:

@misc{xie2025infir,
  title={InfiR: Crafting Effective Small Language Models and Multimodal Small Language Models in Reasoning},
  author={Xie, Congkai and Cai, Shuo and Wang, Wenjun and others},
  year={2025},
  eprint={2502.11573},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}

APA:

Xie, C., Cai, S., Wang, W., et al. (2025). InfiR: Crafting Effective Small Language Models and Multimodal Small Language Models in Reasoning. arXiv:2502.11573.

---

Glossary

• SLM: Small Language Model (<2 B parameters)
• CoT: Chain-of-Thought prompting or training
• REC: Renewable Energy Certificate
• PUE: Power Usage Effectiveness (ratio of total facility power to IT power)

---