Trillion-7B-preview

logo

Introduction

We introduce Trillion-7B-preview, a preview of our latest large language model designed to push the boundaries of multilingual scalability and performance.

When comparing performance to training FLOPs for Trillion-7B-preview with competitive models, our model pushes the Pareto frontier, achieving around 66.5% average performance while using significantly fewer compute (~9.3×10²² FLOPs). It outperforms models like Mistral-7B-Instruct-v0.3 and SOLAR-10.7B-Instruct-v1.0 while remaining competitive with models requiring 3-8× more compute such as Qwen2.5-7B-Instruct and EXAONE-3.5-7.8B-Instruct. For full benchmark results, see tables below.

Average Performance vs. Approximate Training FLOPs

  • Type: Causal Language Model
  • Training Stage: Pre-training & Post-training
  • Architecture: Transformer Decoder with RoPE, SwiGLU, RMSNorm
  • Number of Parameters: 7.76B
  • Number of Layers: 32
  • Number of Attention Heads: 32
  • Context Length: 4,096
  • Number of Tokens seen: 2T
  • Vocab Size: 128,128

Quickstart

Here is a code snippet with apply_chat_template that demonstrates how to load the tokenizer and model and generate text.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "trillionlabs/Trillion-7B-preview"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "Tell me a hilarious knock knock joke."
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    model_inputs["input_ids"],
    attention_mask=model_inputs["attention_mask"],
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

"""
Sure! Here's a classic knock-knock joke that's guaranteed to make you chuckle:
Knock, knock.
Who's there?
Lettuce.
Lettuce who?
Lettuce in, it's too cold out here!
"""

Evaluation

We select a wide variety of benchmarks that evaluate general reasoning, knowledge recall, coding abilities, mathematical reasoning, and instruction following capabilities. We evaluated Trillion-7B-preview along with several leading large language models of similar size. Our model especially demonstrates strong performance on Korean benchmarks.

Full evaluation settings
Benchmark Language Evaluation Setting Metric
General Reasoning and Reading Comprehension
• HellaSwag English 0-shot accuracy
• TruthfulQA_mc1 English 6-shot accuracy
• TruthfulQA_mc2 English 6-shot accuracy
• ARC:C English 0-shot accuracy
• HAERAE Korean 3-shot accuracy
• KoBEST Korean 5-shot accuracy
• BBH English 0-shot, CoT accuracy
• xwinograd_en English 0-shot accuracy
• xwinograd_jp Japanese 0-shot accuracy
• xwinograd_zh Chinese 0-shot accuracy
Knowledge Recall
• KMMLU Korean 5-shot accuracy
• MMLU English 5-shot accuracy
• Global-MMLU-Lite-en English 5-shot accuracy
• Global-MMLU-Lite-ko Korean 5-shot accuracy
• Global-MMLU-Lite-ja Japanese 5-shot accuracy
• Global-MMLU-Lite-zh Chinese 5-shot accuracy
Coding
• HumanEval English 0-shot, CoT pass@1
• MBPP English 0-shot, CoT pass@1
Mathematical Reasoning
• GSM8k English 0-shot, CoT exact-match
• MATH English 0-shot, CoT exact-match
• GPQA English 4-shot accuracy
• HRM8k Korean 0-shot, CoT exact-match
Instruction Following and Chat
• IFEval English 0-shot strict-average
• koIFEval* Korean 0-shot strict-average
• MT-Bench** English LLM-as-a-judge (gpt-4o-2024-08-06) LLM score
• KO-MT-Bench** Korean LLM-as-a-judge (gpt-4o-2024-08-06) LLM score
• LogicKor** Korean LLM-as-a-judge (gpt-4o-2024-08-06) LLM score
  • *Note that koIFEval is our in-house evaluation benchmark for assessing instruction-following capabilities in Korean.
  • **Note that MT-Bench, KO-MT-Bench, and LogicKor use a 10-point scale.

Benchmark Results

General Reasoning and Factuality

Benchmark Trillion-7B-preview EXAONE-3.5-7.8B-Instruct gemma-2-9b-it Llama-3.1-8B-Instruct Qwen2.5-7B-Instruct SOLAR-10.7B-Instruct-v1.0 Mistral-7B-Instruct-v0.3
HellaSwag 58.94 60.04 59.72 59.81 61.97 68.72 65.79
TruthfulQA_mc1 36.10 40.64 42.96 38.07 47.74 56.18 42.47
TruthfulQA_mc2 54.10 59.74 60.09 54.54 64.72 70.64 59.41
ARC:C 54.44 56.40 62.97 53.58 52.99 60.07 58.11
HAERAE 80.02 76.08 68.01 63.15 65.17 60.86 47.75
KoBEST 79.61 78.57 79.98 70.09 79.24 75.20 66.50
KMMLU 48.09 45.39 46.66 41.41 50.15 41.66 33.59
MMLU 63.52 65.65 72.24 68.32 74.23 65.20 61.84
Global-MMLU-Lite-en 67.75 69.50 76.25 67.50 77.25 71.75 65.50
Global-MMLU-Lite-ko 60.75 60.00 64.25 54.00 59.25 53.75 43.00
Global-MMLU-Lite-ja 60.75 45.75 66.50 54.50 65.75 50.75 50.00
Global-MMLU-Lite-zh 59.50 50.00 63.75 60.25 68.75 57.00 47.25
BBH 41.94 53.30 28.77 43.16 53.68 52.91 45.09
xwinograd_en 87.78 87.10 89.55 88.09 85.63 87.35 88.39
xwinograd_jp 79.98 74.45 80.92 76.02 72.89 72.58 70.70
xwinograd_zh 73.81 69.44 68.06 76.19 81.55 74.60 71.83

Coding

Benchmark Trillion-7B-preview EXAONE-3.5-7.8B-Instruct gemma-2-9b-it Llama-3.1-8B-Instruct Qwen2.5-7B-Instruct SOLAR-10.7B-Instruct-v1.0 Mistral-7B-Instruct-v0.3
HumanEval 55.48 79.26 60.98 67.68 81.71 34.76 36.59
MBPP 40.40 61.40 8.40 39.20 51.00 29.40 36.00

Mathematical Reasoning

Benchmark Trillion-7B-preview EXAONE-3.5-7.8B-Instruct gemma-2-9b-it Llama-3.1-8B-Instruct Qwen2.5-7B-Instruct SOLAR-10.7B-Instruct-v1.0 Mistral-7B-Instruct-v0.3
GSM8k 72.25 87.79 73.69 74.98 88.86 62.93 35.94
MATH 32.70 70.68 - 38.30 71.50 14.38 12.12
GPQA 32.81 38.61 36.83 30.58 34.15 28.35 32.59
HRM8k 30.10 38.99 16.04 - 41.51 20.68 7.89

Instruction Following and Chat

Benchmark Trillion-7B-preview EXAONE-3.5-7.8B-Instruct gemma-2-9b-it Llama-3.1-8B-Instruct Qwen2.5-7B-Instruct SOLAR-10.7B-Instruct-v1.0 Mistral-7B-Instruct-v0.3
IFEval 79.13 81.42 75.48 74.93 75.85 51.61 52.64
koIFEval 66.58 54.65 43.30 36.07 48.55 26.12 34.22
MT-Bench 6.53 6.75 - 6.32 7.86 6.76 6.84
KO-MT-Bench 6.21 6.70 - 4.27 6.47 5.57 4.59
LogicKor 8.14 9.25 8.33 6.45 7.99 1.85 4.76

Limitations

  • Language Support: The model is optimized for English, Korean, Japanese, and Chinese. Usage with other languages may result in degraded performance.
  • Knowledge Cutoff: The model's information is limited to data available up to August 2023.
  • Safety Mechanisms: This release does not yet include comprehensive safety features. Future updates will address this area.
  • Release Status: This is a preliminary release version with planned enhancements and updates forthcoming.

License

This model repository is licensed under the Apache-2.0 License.

Citation

@article{trillion7Bpreview,
  title={Trillion-7B-preview},
  author={trillionlabs},
  year={2025},
  url={https://huggingface.co/trillionlabs/Trillion-7B-preview}
}

Contact

For inquiries, please contact: [email protected]

Downloads last month
1,603
Safetensors
Model size
7.53B params
Tensor type
BF16
·
Inference Providers NEW
This model is not currently available via any of the supported Inference Providers.