SeedGPT-V3 is a fine-tuned causal language model (chat-style LLM) based on the custom base model SeedGPT-V2. This version is optimized for dialogue and assistant-style conversations and trained on the lmsys-chat-1m dataset, which contains high-quality human-chat interaction data.

The model follows a Transformer-based architecture and is compatible with 🤗 Hugging Face's AutoModelForCausalLM interface using a custom configuration and model class.

Developed by: Sumon Singh
Model type: Causal Language Model (Chat-optimized Transformer)
Language(s): English
License: MIT
Finetuned from model: SeedGPT-V2

Model Sources

Dataset: lmsys-chat-1m
Repository: GitHub - SeedGPT-22M
Demo: Hugging Face Space

Uses

SeedGPT-V3 is intended for chat-style text generation, including:

🤖 Conversational AI assistants
✍️ Creative writing in dialogue form
💬 Roleplay bots or dialogue simulations

Out-of-Scope Use

❌ Not suitable for high-stakes decision-making (e.g., legal, medical, financial)
❌ Not designed for factual QA systems or real-time reliability
❌ Avoid using for advice without human validation

Bias, Risks, and Limitations

As with most large language models, SeedGPT-V3 may:

🔄 Hallucinate or fabricate information
⚠️ Exhibit biases present in training data (gender, race, etc.)
🧠 Lack factual grounding or real-world knowledge after 2023

Recommendations

Always validate generated content before relying on it. For safety-critical applications, do not use without additional alignment, filtering, or oversight.

How to Get Started with the Model

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from transformers.models.auto.configuration_auto import CONFIG_MAPPING
from transformers.models.auto.modeling_auto import MODEL_FOR_CAUSAL_LM_MAPPING
from src.convert_to_hf_model import HFTransformerConfig, HFTransformerModel  # custom config and model
import torch

# Register custom config and model
CONFIG_MAPPING.register("hf_transformer", HFTransformerConfig)
MODEL_FOR_CAUSAL_LM_MAPPING.register(HFTransformerConfig, HFTransformerModel)

# Load tokenizer and model

tokenizer = AutoTokenizer.from_pretrained("singhsumony2j/SeedGPT-V3")
model = AutoModelForCausalLM.from_pretrained("singhsumony2j/SeedGPT-V3", low_cpu_mem_usage=True)

# Apply style template

tokenizer.chat_template = """
{% for message in messages %}
{% if message["role"] == "user" %}
<S>user: {{ message["content"] }}</S>
{% elif message["role"] == "assistant" %}
<S>assistant: {{ message["content"] }}</S>
{% endif %}
{% endfor %}
{% if add_generation_prompt %}
<S>assistant:
{% endif %}
"""

#Example chat prompt using the tokenizer's chat template

prompt = "Hi, how are you?"
chat = [{"role": "user", "content": prompt}]

input_txt = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(input_txt, return_tensors="pt")

# Generate response
output = model.generate(inputs["input_ids"], max_tokens=max_num_tokens, temp=temp)
generated = output[0][inputs["input_ids"].shape[1]:]
output = tokenizer.decode(generated, skip_special_tokens=True)
print(output)

Citation

bibtex

@misc{singh2025seedgptv3,
  author = {Sumon Singh},
  title = {SeedGPT-V3: A Fine-Tuned Chat Language Model based on SeedGPT-V2},
  year = {2025},
  howpublished = {\url{https://huggingface.co/singhsumony2j/SeedGPT-V3}}
}

Model Card Authors

Sumon Singh

Model Card Contact

singhsumony2j@gmail.com