---
language:
- en
- es
- fr
- de
- it
- pt
- nl
- ru
- zh
- ja
- ko
- ar
- hi
license: apache-2.0
library_name: transformers
tags:
- text-generation
- conversational
- code
- instruction-following
- pytorch
- causal-lm
- llm
- reasoning
- multilingual
pipeline_tag: text-generation
widget:
- text: "def fibonacci(n):"
example_title: Code Generation
- text: "Explain quantum entanglement in simple terms:"
example_title: Science Explanation
- text: "Write a short story about a robot learning to paint:"
example_title: Creative Writing
model-index:
- name: Helion-V2
results:
- task:
type: text-generation
name: Text Generation
dataset:
name: MMLU
type: cais/mmlu
metrics:
- type: accuracy
value: 64.2
name: Accuracy
- task:
type: text-generation
name: Code Generation
dataset:
name: HumanEval
type: openai_humaneval
metrics:
- type: pass@1
value: 48.2
name: Pass@1
- task:
type: text-generation
name: Commonsense Reasoning
dataset:
name: HellaSwag
type: hellaswag
metrics:
- type: acc_norm
value: 80.5
name: Accuracy
- task:
type: text-generation
name: Truthfulness
dataset:
name: TruthfulQA
type: truthful_qa
metrics:
- type: mc2
value: 52.1
name: MC2
- task:
type: text-generation
name: Math Reasoning
dataset:
name: GSM8K
type: gsm8k
metrics:
- type: accuracy
value: 68.7
name: Accuracy
- task:
type: text-generation
name: Question Answering
dataset:
name: ARC Challenge
type: ai2_arc
metrics:
- type: acc_norm
value: 58.3
name: Accuracy
---
# Helion-V2
---
**A State-of-the-Art 7.2B Parameter Language Model for Daily Use**
[](https://opensource.org/licenses/Apache-2.0)
[](https://www.python.org/downloads/)
[](https://github.com/huggingface/transformers)
[](https://pytorch.org/)
[Model Card](#model-information) | [Usage](#usage) | [Benchmarks](#performance-benchmarks) | [Safety](#safety-and-moderation)
---
## Table of Contents
- [Model Overview](#model-overview)
- [Model Information](#model-information)
- [Performance Benchmarks](#performance-benchmarks)
- [Quick Start](#quick-start)
- [Usage](#usage)
- [Safety and Moderation](#safety-and-moderation)
- [Deployment Options](#deployment-options)
- [Training Details](#training-details)
- [Limitations](#limitations)
- [Citation](#citation)
- [License](#license)
---
## Model Overview
Helion-V2 is an advanced large language model engineered for practical, everyday applications. With 7.2 billion parameters and a focus on factual accuracy, conversational ability, and code generation, Helion-V2 delivers enterprise-grade performance on consumer hardware.
**Key Highlights:**
- **7.2B parameters** optimized for efficiency and quality
- **8,192 token context** for handling complex documents
- **Grouped Query Attention (GQA)** for 40% faster inference
- **Exceptional truthfulness** (52.1% on TruthfulQA - highest in class)
- **Strong coding ability** (48.2% on HumanEval)
- **Multi-language support** with primary focus on English
- **Apache 2.0 License** for commercial use
---
## Model Information
### Architecture Details
| Specification | Value |
|--------------|-------|
| **Parameters** | 7.2 billion |
| **Architecture** | Decoder-only Transformer |
| **Layers** | 32 |
| **Hidden Dimension** | 4,096 |
| **Attention Heads** | 32 (query) / 8 (key-value) |
| **FFN Dimension** | 14,336 |
| **Context Length** | 8,192 tokens |
| **Vocabulary Size** | 32,768 tokens |
| **Position Encoding** | RoPE (Rotary Position Embedding) |
| **Normalization** | RMSNorm (eps: 1e-6) |
| **Activation** | SiLU (Swish) |
| **Attention Type** | Grouped Query Attention (GQA) |
### Model Card Metadata
| Property | Details |
|----------|---------|
| **Model Type** | Causal Language Model |
| **Languages** | English (primary), Spanish, French, German, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, Korean, Arabic, Hindi |
| **License** | Apache 2.0 |
| **Training Data** | 2.5T tokens (web, code, books, papers) |
| **Knowledge Cutoff** | October 2024 |
| **Developed By** | DeepXR |
| **Model Family** | Helion |
| **Version** | 2.0 |
| **Release Date** | November 2024 |
| **Precision** | BFloat16 / Float16 |
| **Framework** | PyTorch 2.1+ |
| **Compute Type** | GPU (NVIDIA A100, H100, RTX 4090+) |
| **Finetuned From** | Trained from scratch |
| **Training Duration** | 21 days on 128x H100 GPUs |
### Supported Tasks
- **Text Generation**: Articles, stories, essays, reports
- **Conversational AI**: Multi-turn dialogue, chat applications
- **Code Generation**: Python, JavaScript, Java, C++, and 20+ languages
- **Question Answering**: Factual queries, reasoning tasks
- **Text Summarization**: Document condensation, key point extraction
- **Creative Writing**: Storytelling, poetry, scriptwriting
- **Data Analysis**: Interpretation, insights, recommendations
- **Translation**: 13 language pairs (quality varies)
- **Educational Tutoring**: Math, science, history, programming
- **Business Writing**: Emails, proposals, presentations
---
## Performance Benchmarks
### Comprehensive Evaluation Results
Helion-V2 has been evaluated on 15+ industry-standard benchmarks, demonstrating strong performance across reasoning, knowledge, coding, and safety metrics.
#### Core Academic Benchmarks
| Benchmark | Helion-V2 | Llama-3-8B | Mistral-7B-v0.3 | Gemma-7B | Qwen-2-7B | GPT-3.5-Turbo |
|-----------|-----------|------------|-----------------|----------|-----------|---------------|
| **MMLU** (5-shot) | **64.2** | 66.4 | 62.5 | 64.3 | 65.1 | 70.0 |
| **MMLU-Pro** (5-shot) | **41.8** | 43.2 | 38.6 | 40.1 | 42.3 | 48.5 |
| **HellaSwag** (10-shot) | **80.5** | 82.1 | 81.3 | 80.9 | 81.7 | 85.5 |
| **PIQA** (0-shot) | **79.8** | 80.5 | 79.1 | 79.6 | 80.2 | 81.6 |
| **WinoGrande** (5-shot) | **74.3** | 75.1 | 73.2 | 74.0 | 74.8 | 77.2 |
| **ARC-Challenge** (25-shot) | **58.3** | 59.2 | 56.7 | 57.9 | 58.8 | 61.4 |
| **ARC-Easy** (25-shot) | **82.7** | 83.4 | 81.9 | 82.5 | 83.1 | 85.2 |
| **OpenBookQA** (10-shot) | **51.6** | 52.8 | 49.4 | 50.9 | 52.1 | 54.3 |
#### Mathematical and Logical Reasoning
| Benchmark | Helion-V2 | Llama-3-8B | Mistral-7B-v0.3 | Gemma-7B | Qwen-2-7B | GPT-3.5-Turbo |
|-----------|-----------|------------|-----------------|----------|-----------|---------------|
| **GSM8K** (8-shot CoT) | **68.7** | 72.4 | 52.3 | 66.1 | 71.8 | 77.3 |
| **MATH** (4-shot) | **23.5** | 26.8 | 15.2 | 21.7 | 25.4 | 34.1 |
| **BBH** (3-shot) | **52.9** | 55.3 | 49.1 | 51.6 | 54.2 | 60.7 |
| **DROP** (3-shot) | **61.4** | 63.7 | 58.2 | 60.5 | 62.8 | 68.3 |
#### Code Generation and Understanding
| Benchmark | Helion-V2 | Llama-3-8B | Mistral-7B-v0.3 | Gemma-7B | Qwen-2-7B | CodeLlama-7B |
|-----------|-----------|------------|-----------------|----------|-----------|--------------|
| **HumanEval** (pass@1) | **48.2** | 51.8 | 40.2 | 44.5 | 49.7 | 45.9 |
| **HumanEval** (pass@10) | **67.3** | 71.2 | 59.8 | 64.1 | 68.9 | 66.2 |
| **MBPP** (pass@1) | **55.8** | 58.3 | 47.1 | 52.6 | 57.4 | 54.1 |
| **MBPP** (pass@10) | **74.6** | 77.9 | 68.3 | 72.1 | 76.2 | 73.8 |
| **MultiPL-E** (Python) | **46.9** | 49.5 | 38.7 | 43.2 | 48.1 | 44.6 |
| **MultiPL-E** (JavaScript) | **43.5** | 46.2 | 35.9 | 40.8 | 44.7 | 41.3 |
| **DS-1000** (Data Science) | **38.7** | 41.2 | 32.4 | 36.9 | 40.3 | 37.5 |
#### Truthfulness and Safety
| Benchmark | Helion-V2 | Llama-3-8B | Mistral-7B-v0.3 | Gemma-7B | Qwen-2-7B | GPT-3.5-Turbo |
|-----------|-----------|------------|-----------------|----------|-----------|---------------|
| **TruthfulQA** (MC2) | **52.1** | 48.3 | 47.6 | 49.2 | 51.3 | 54.7 |
| **TruthfulQA** (MC1) | **37.8** | 34.6 | 33.9 | 35.7 | 37.1 | 40.2 |
| **ToxiGen** (lower is better) | **0.08** | 0.12 | 0.15 | 0.10 | 0.09 | 0.06 |
| **CrowS-Pairs** (bias score) | **54.2** | 57.8 | 59.3 | 56.1 | 55.0 | 52.1 |
#### Conversational and Instruction Following
| Benchmark | Helion-V2 | Llama-3-8B | Mistral-7B-v0.3 | Gemma-7B | Qwen-2-7B | GPT-3.5-Turbo |
|-----------|-----------|------------|-----------------|----------|-----------|---------------|
| **MT-Bench** (Avg) | **7.85** | 8.12 | 7.61 | 7.73 | 7.92 | 8.32 |
| **AlpacaEval 2.0** (Win Rate) | **18.3%** | 22.1% | 14.7% | 16.8% | 19.4% | 28.5% |
| **Arena-Hard** | **31.7** | 35.4 | 27.8 | 29.9 | 33.2 | 42.6 |
| **IFEval** (Instruction Following) | **72.4** | 75.8 | 68.9 | 71.2 | 74.1 | 78.3 |
### Performance Analysis
**Strengths:**
- **Truthfulness Leader**: Highest TruthfulQA score in its parameter class (52.1%), demonstrating superior factual accuracy and reduced hallucination
- **Safety-First Design**: Lowest toxicity score (0.08 on ToxiGen) and competitive bias metrics
- **Balanced Capabilities**: Strong performance across all task categories without extreme specialization
- **Code Competence**: 48.2% HumanEval pass@1 places it among top general-purpose 7B models
- **Practical Focus**: Optimized for real-world use cases rather than benchmark gaming
**Comparative Advantages:**
- 8% more truthful than Llama-3-8B on TruthfulQA
- 33% less toxic than Mistral-7B-v0.3 on ToxiGen
- Better instruction following than Gemma-7B on IFEval
- More balanced than specialized models (e.g., better general knowledge than CodeLlama)
**Areas for Improvement:**
- Math performance trails Llama-3-8B and Qwen-2-7B by ~4-5%
- Conversational win rate below top performers on AlpacaEval 2.0
- Complex reasoning (BBH, MATH) shows room for enhancement
### Inference Performance
| Configuration | Hardware | Throughput | Latency (TTFT) | Memory |
|---------------|----------|------------|----------------|--------|
| FP16 | A100 (80GB) | 52 tokens/s | 87ms | 14.4 GB |
| FP16 | RTX 4090 (24GB) | 47 tokens/s | 102ms | 14.4 GB |
| 8-bit | RTX 4090 (24GB) | 41 tokens/s | 115ms | 7.8 GB |
| 4-bit | RTX 3090 (24GB) | 38 tokens/s | 128ms | 4.2 GB |
| 4-bit | RTX 3060 (12GB) | 29 tokens/s | 156ms | 4.2 GB |
*TTFT = Time To First Token; Measured with 2048 token context, 512 token generation*
---
## Quick Start
### Installation
```bash
pip install transformers torch accelerate bitsandbytes safetensors
```
### Basic Usage
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_name = "DeepXR/Helion-V2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.float16,
device_map="auto"
)
prompt = "Explain the theory of relativity in simple terms:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
**inputs,
max_new_tokens=256,
temperature=0.7,
top_p=0.9,
do_sample=True
)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
```
---
## Usage
### Chat Interface
```python
messages = [
{"role": "system", "content": "You are a helpful, respectful, and honest AI assistant."},
{"role": "user", "content": "Write a Python function to calculate fibonacci numbers."}
]
input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
outputs = model.generate(
**inputs,
max_new_tokens=512,
temperature=0.7,
top_p=0.9,
repetition_penalty=1.1
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```
### Advanced Generation Parameters
```python
# For creative writing
outputs = model.generate(
**inputs,
max_new_tokens=1024,
temperature=0.9,
top_p=0.95,
top_k=50,
repetition_penalty=1.15
)
# For factual/technical content
outputs = model.generate(
**inputs,
max_new_tokens=512,
temperature=0.3,
top_p=0.85,
repetition_penalty=1.05
)
# For code generation
outputs = model.generate(
**inputs,
max_new_tokens=1024,
temperature=0.2,
top_p=0.9,
repetition_penalty=1.1
)
```
### Quantization for Efficient Deployment
#### 4-bit Quantization (Recommended)
```python
from transformers import BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4"
)
model = AutoModelForCausalLM.from_pretrained(
"DeepXR/Helion-V2",
quantization_config=quantization_config,
device_map="auto"
)
```
#### 8-bit Quantization
```python
model = AutoModelForCausalLM.from_pretrained(
"DeepXR/Helion-V2",
load_in_8bit=True,
device_map="auto"
)
```
### Streaming Generation
```python
from transformers import TextIteratorStreamer
from threading import Thread
streamer = TextIteratorStreamer(tokenizer, skip_special_tokens=True)
generation_kwargs = dict(
inputs,
streamer=streamer,
max_new_tokens=512,
temperature=0.7,
top_p=0.9
)
thread = Thread(target=model.generate, kwargs=generation_kwargs)
thread.start()
for new_text in streamer:
print(new_text, end="", flush=True)
```
---
## Safety and Moderation
Helion-V2 incorporates multiple safety layers to ensure responsible AI deployment:
### Built-in Safety Features
1. **Content Filtering**: Training data filtered for toxicity, hate speech, and explicit content
2. **Bias Mitigation**: Balanced representation across demographics and viewpoints
3. **Truthfulness Optimization**: Enhanced training to reduce hallucinations
4. **Instruction Compliance**: Fine-tuned to decline harmful requests appropriately
### Safety Scores
- **ToxiGen Score**: 0.08 (Lower is better; competitive with GPT-3.5)
- **CrowS-Pairs Bias**: 54.2 (Near-neutral; 50 is perfect balance)
- **TruthfulQA**: 52.1% (Highest in 7B parameter class)
- **RealToxicityPrompts**: 2.1% toxic completions (with default sampling)
### Recommended Safety Measures
For production deployments, we recommend implementing:
1. **Content Moderation API**: Use the provided `safety_classifier.py` for output filtering
2. **Input Validation**: Screen user inputs for malicious prompts
3. **Rate Limiting**: Prevent abuse through usage caps
4. **Monitoring**: Log and review model interactions
5. **Human Oversight**: Implement human-in-the-loop for sensitive applications
### Using the Safety Classifier
```python
from safety_classifier import SafetyClassifier
safety = SafetyClassifier()
# Check if prompt is safe
is_safe, category = safety.check_prompt(user_input)
if not is_safe:
print(f"Unsafe prompt detected: {category}")
# Handle appropriately
# Check model output
response = model.generate(...)
is_safe, category = safety.check_response(response)
if not is_safe:
# Filter or regenerate response
response = safety.sanitize_response(response)
```
See `safety_classifier.py` and `content_moderation.py` for complete implementation.
---
## Deployment Options
### Local Deployment
**Recommended Hardware:**
- GPU: NVIDIA RTX 3090/4090 (24GB) or better
- RAM: 32GB+ system memory
- Storage: 20GB for model files
### Cloud Deployment
**Optimized Configurations:**
```python
# AWS SageMaker
from sagemaker.huggingface import HuggingFaceModel
huggingface_model = HuggingFaceModel(
model_data="s3://your-bucket/helion-v2",
role=role,
transformers_version="4.40",
pytorch_version="2.1",
py_version="py310",
)
predictor = huggingface_model.deploy(
initial_instance_count=1,
instance_type="ml.g5.2xlarge"
)
```
### API Server
```python
# Using FastAPI
from fastapi import FastAPI
from pydantic import BaseModel
app = FastAPI()
class GenerationRequest(BaseModel):
prompt: str
max_tokens: int = 256
temperature: float = 0.7
@app.post("/generate")
async def generate(request: GenerationRequest):
inputs = tokenizer(request.prompt, return_tensors="pt").to(device)
outputs = model.generate(
**inputs,
max_new_tokens=request.max_tokens,
temperature=request.temperature
)
return {"response": tokenizer.decode(outputs[0], skip_special_tokens=True)}
```
### GGUF Format (llama.cpp)
For CPU inference and edge deployment:
```bash
# Download GGUF quantized version
wget https://huggingface.co/DeepXR/Helion-V2-GGUF/resolve/main/helion-v2-q4_k_m.gguf
# Run with llama.cpp
./llama-cli -m helion-v2-q4_k_m.gguf -p "Your prompt here" -n 256
```
---
## Training Details
### Training Data Composition
| Data Source | Percentage | Tokens | Description |
|------------|------------|--------|-------------|
| Web Documents | 45% | 1.125T | High-quality web pages, articles, documentation |
| Code Repositories | 20% | 500B | GitHub, Stack Overflow, technical forums |
| Books | 15% | 375B | Fiction, non-fiction, educational materials |
| Scientific Papers | 10% | 250B | ArXiv, PubMed, academic publications |
| Instruction Data | 10% | 250B | Curated instruction-response pairs |
**Total Training Tokens**: 2.5 trillion
### Data Processing Pipeline
1. **Collection**: Scraped from verified sources with license compliance
2. **Quality Filtering**: Perplexity-based filtering (threshold: 2000)
3. **Deduplication**: MinHash LSH for near-duplicate removal (>95% similarity)
4. **Toxicity Filtering**: Removed content flagged by Perspective API (score >0.7)
5. **PII Removal**: Named entity recognition and regex-based scrubbing
6. **Language Detection**: Filtered for 13 target languages
7. **Code Quality**: AST validation, syntax checking, license verification
### Training Hyperparameters
| Parameter | Value |
|-----------|-------|
| Optimizer | AdamW |
| Peak Learning Rate | 3e-4 |
| Learning Rate Schedule | Cosine with warmup |
| Warmup Steps | 2,000 |
| Weight Decay | 0.01 |
| Gradient Clipping | 1.0 |
| Batch Size | 4M tokens |
| Sequence Length | 8,192 tokens |
| Training Steps | 600,000 |
| Epochs | 3 |
| Precision | BFloat16 |
| Beta1 | 0.9 |
| Beta2 | 0.95 |
| Epsilon | 1e-8 |
### Infrastructure
- **GPUs**: 128x NVIDIA H100 80GB (SXM5)
- **Framework**: PyTorch 2.1.2 with CUDA 12.1
- **Distributed Training**: DeepSpeed ZeRO-3 with CPU offloading
- **Mixed Precision**: BFloat16 with gradient scaling
- **Checkpointing**: Every 1,000 steps (3 checkpoints retained)
- **Training Duration**: 21 days
- **Total GPU Hours**: 64,512 hours
- **Estimated Cost**: $450,000 USD
### Post-Training Refinement
1. **Supervised Fine-Tuning (SFT)**: 150,000 instruction-response pairs
2. **Direct Preference Optimization (DPO)**: 50,000 preference pairs
3. **Safety Fine-Tuning**: 25,000 safety-focused examples
4. **Evaluation-Driven Refinement**: Iterative improvements based on benchmark performance
---
## Limitations
### Known Limitations
1. **Temporal Knowledge**: Information cutoff at October 2024; no awareness of events after this date
2. **Hallucination Risk**: May generate plausible but incorrect information (mitigated but not eliminated)
3. **Context Length**: Performance degrades beyond 6,000 tokens despite 8,192 token capacity
4. **Mathematical Reasoning**: Struggles with complex multi-step calculations requiring precise arithmetic
5. **Specialized Domains**: Limited accuracy in highly technical fields (e.g., advanced physics, medicine, law)
6. **Language Imbalance**: Best performance in English; variable quality in other languages
7. **Code Debugging**: Better at generation than debugging complex existing codebases
8. **Long-Term Memory**: No persistent memory across conversations
9. **Real-Time Information**: Cannot access current data, news, or live information
10. **Multimodal Understanding**: Text-only model; no image, audio, or video processing
### Ethical Considerations
**Bias**: Training data may reflect societal biases related to gender, race, culture, geography, and socioeconomic status. Users should validate outputs for fairness.
**Misuse Potential**: Model can be misused for generating misinformation, spam, or harmful content. Implement appropriate safeguards.
**Environmental Impact**: Training consumed significant energy (est. 8,500 kg CO2eq). Consider carbon offset for large-scale deployments.
**Privacy**: Do not input personally identifiable information (PII) or confidential data without encryption and proper handling.
### Use Case Restrictions
**DO NOT USE FOR:**
- Medical diagnosis or treatment recommendations
- Legal advice or contractual interpretation
- Financial investment decisions
- Safety-critical systems (aviation, automotive, medical devices)
- Autonomous decision-making without human oversight
- Generating false identification or credentials
- Impersonating individuals or organizations
- Processing sensitive personal data without consent
---
## Citation
If you use Helion-V2 in your research or applications, please cite:
```bibtex
@misc{helion-v2-2024,
title={Helion-V2: An Efficient and Truthful Large Language Model for Daily Use},
author={DeepXR Team},
year={2025},
month={November},
publisher={HuggingFace},
url={https://huggingface.co/DeepXR/Helion-V2},
note={7.2B parameter decoder-only transformer with grouped query attention}
}
```
For technical details:
```bibtex
@techreport{helion-v2-technical-2025,
title={Helion-V2: Technical Report},
author={DeepXR Research Team},
institution={DeepXR},
year={2025},
type={Technical Report},
url={https://deepxr.ai/research/helion-v2-technical-report.pdf}
}
```
---
## License
This model is released under the **Apache License 2.0**. You are free to:
- Use commercially
- Modify and distribute
- Use privately
- Use for patent purposes
**Conditions:**
- Include copyright notice
- Include license copy
- State changes made
- Include NOTICE file if present
See [LICENSE](LICENSE) file for complete terms.
---
## Acknowledgments
We extend our gratitude to:
- **Hugging Face** for the Transformers library and model hosting infrastructure
- **PyTorch Team** for the deep learning framework
- **DeepSpeed Team** (Microsoft) for distributed training tools
- **EleutherAI** for evaluation frameworks and benchmarks
- **Open Source Community** for datasets, tools, and collaborative research
- **Our Compute Partners** for providing GPU infrastructure
Special thanks to researchers whose work influenced this project: LLaMA, Mistral, GPT, PaLM, and countless others advancing open language models.
---
**Developed with care by the DeepXR Team**
*Building responsible, capable, and accessible AI for everyone*
