Model Summary:
Granite-4.0-H-Small-Base is a decoder-only, long-context language model designed for a wide range of text-to-text generation tasks. It also supports Fill-in-the-Middle (FIM) code completion through the use of specialized prefix and suffix tokens. The model is trained from scratch on approximately 23 trillion tokens following a four-stage training strategy: 15 trillion tokens in the first stage, 5 trillion in the second, 2 trillion in the third, and 0.5 trillion in the final stage.
Supported Languages:
English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese. Users may finetune Granite 4.0 models for languages beyond these languages.
Intended Use:
Prominent use cases of LLMs in text-to-text generation include summarization, text classification, extraction, question-answering, code-completion (including FIM), and long-context generation tasks. All Granite Base models are able to handle these tasks as they were trained on a large amount of data from various domains. Moreover, they can serve as baseline to create specialized models for specific application scenarios.
Generation:
This is a simple example of how to use Granite-4.0-H-Small-Base model.
Then, copy the code snippet below to run the example.
python
1from transformers import AutoModelForCausalLM, AutoTokenizer
2device ="cuda"34model_path ="ibm-granite/granite-4.0-h-micro-base"56tokenizer = AutoTokenizer.from_pretrained(model_path)7# drop device_map if running on CPU8model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device)9model.eval()10# change input text as desired11input_text ="The capital of France is"12# tokenize the text13input_tokens = tokenizer(input_text, return_tensors="pt").to(device)14# generate output tokens15output = model.generate(**input_tokens, max_length=10)16# decode output tokens into text17output = tokenizer.batch_decode(output)18# print output19print(output[0])
Expected output:
The capital of France is Paris.
Evaluation Results:
Benchmarks
Metric
Micro Dense
H Micro Dense
H Tiny MoE
H Small MoE
General Tasks
MMLU
5-shot
66.47
67.43
68.90
75.85
MMLU-Pro
5-shot,CoT
37.16
34.03
35.47
48.94
BBH
3-shot, CoT
63.84
57.65
59.67
75.84
AGI EVAL
3-shot
54.32
54.59
53.69
62.05
DROP
5-shot
66.04
67.44
64.92
74.69
Math Tasks
GSM8K
8-shot
72.93
63.76
72.55
82.11
Minerva Math
4-shot
38
39.7
40.34
46.28
Code Tasks
HumanEval
pass@1 [StarCoder Prompt]
76.19
73.72
77.59
83.66
HumanEval
pass@1
59.76
70.73
71.34
76.22
HumanEval+
pass@1
54.27
67.07
64.02
69.51
MBPP
pass@1
81.48
74.87
81.48
83.07
MBPP+
pass@1
68.25
63.23
68.78
70.37
Multilingual Tasks
MMMLU
5-shot
56.59
58.5
62.77
71.18
INCLUDE
5-shot
51.77
52.16
53.78
66.04
MGSM
8-shot
58.48
47.04
54.64
65.2
Multilingual Benchmarks and thr included languages:
Benchmarks
# Langs
Languages
MMMLU
11
ar, de, en, es, fr, ja, ko, pt, zh, bn, hi
INCLUDE
14
hi, bn, ta, te, ar, de, es, fr, it, ja, ko, nl, pt, zh
MGSM
5
en, es, fr, ja, zh
Model Architecture:
Granite-4.0-H-Small-Base is based on a decoder-only MoE transformer architecture. Core components of this architecture are: GQA, Mamba2, MoEs with shared experts, SwiGLU activation, RMSNorm, and shared input/output embeddings.
Model
Micro Dense
H Micro Dense
H Tiny MoE
H Small MoE
Embedding size
2560
2048
1536
4096
Number of layers
40 attention
4 attention / 36 Mamba2
4 attention / 36 Mamba2
4 attention / 36 Mamba2
Attention head size
64
64
128
128
Number of attention heads
40
32
12
32
Number of KV heads
8
8
4
8
Mamba2 state size
-
128
128
128
Number of Mamba2 heads
-
64
48
128
MLP / Shared expert hidden size
8192
8192
1024
1536
Num. Experts
-
-
64
72
Num. active Experts
-
-
6
10
Expert hidden size
-
-
512
768
MLP activation
SwiGLU
SwiGLU
SwiGLU
SwiGLU
Sequence length
128K
128K
128K
128K
Position embedding
RoPE
NoPE
NoPE
NoPE
# Parameters
3B
3B
7B
32B
# Active parameters
3B
3B
1B
9B
Training Data: This model is trained on a mix of open source and proprietary data following a four-stage training strategy.
Stage
Characteristics
Micro Dense
H Micro Dense
H Tiny MoE
H Small MoE
I
General mixture of training data, warmup, and power scheduler for learning rate.
10
10
15
15
II
General mixture of training data with higher percentages of code and math with power scheduler for learning rate.
2
5
5
5
III
High quality training data, exponential decay of learning rate.
2
2
2
2
IV
High quality training data, linear decay to zero for learning rate.
0.5
0.5
0.5
0.5
Infrastructure:
We trained the Granite 4.0 Language Models utilizing an NVIDIA GB200 NVL72 cluster hosted in CoreWeave. Intra-rack communication occurs via the 72-GPU NVLink domain, and a non-blocking, full Fat-Tree NDR 400 Gb/s InfiniBand network provides inter-rack communication. This cluster provides a scalable and efficient infrastructure for training our models over thousands of GPUs.
Ethical Considerations and Limitations:
The use of Large Language Models involves risks and ethical considerations people must be aware of, including but not limited to: bias and fairness, misinformation, and autonomous decision-making. Granite-4.0-H-Small-Base model is not the exception in this regard. Even though this model is suited for multiple generative AI tasks, it has not undergone any safety alignment, there it may produce problematic outputs. Additionally, it remains uncertain whether smaller models might exhibit increased susceptibility to hallucination in generation scenarios by copying text verbatim from the training dataset due to their reduced sizes and memorization capacities. This aspect is currently an active area of research, and we anticipate more rigorous exploration, comprehension, and mitigations in this domain. Regarding ethics, a latent risk associated with all Large Language Models is their malicious utilization. We urge the community to use Granite-4.0-H-Small-Base model with ethical intentions and in a responsible way.
