LwQ-Reasoner-10B (Llama with Questions), based on the Llama 3.1 collection of multilingual large language models (LLMs), is a set of pre-trained and instruction-tuned generative models optimized for multilingual dialogue use cases. These models outperform many available open-source alternatives. Model Architecture: Llama 3.1 is an auto-regressive language model utilizing an optimized transformer architecture. The tuned versions undergo supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to better align with human preferences for helpfulness and safety. LwQ-Reasoner-10B is trained on synthetic reasoning datasets for mathematical reasoning and context-based problem-solving, with a focus on following instructions or keywords embedded in the input.

Use with transformers

Starting with transformers >= 4.43.0 onward, you can run conversational inference using the Transformers pipeline abstraction or by leveraging the Auto classes with the generate() function.

Make sure to update your transformers installation via pip install --upgrade transformers.

python

import transformers
import torch

model_id = "prithivMLmods/LwQ-Reasoner-10B"

pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
    {"role": "user", "content": "Who are you?"},
]

outputs = pipeline(
    messages,
    max_new_tokens=256,
)
print(outputs[0]["generated_text"][-1])

Config and Base

json

{
  "_name_or_path": "prithivMLmods/Triangulum-10B",
  "architectures": [
    "LlamaForCausalLM"
  ]
}

Intended Use

Multilingual Dialogue Systems: LwQ-Reasoner-10B is designed for creating conversational agents capable of engaging in dialogues across multiple languages, making it suitable for global customer support and multilingual chatbots.
Instruction-Following Tasks: The model excels at tasks requiring adherence to specific instructions or keywords embedded in the input, such as form completion, task automation, and guided workflows.
Mathematical Reasoning: With specialized training on synthetic reasoning datasets, LwQ-Reasoner-10B can perform complex mathematical reasoning and problem-solving, making it useful for educational platforms, tutoring systems, and research assistance.
Context-Based Problem Solving: The model is optimized to handle contextually rich problems, allowing it to generate context-aware responses for applications such as summarization, question answering, and decision support.
Content Generation: It can generate high-quality content, including articles, reports, summaries, and creative writing, across various domains and languages.
Knowledge Retrieval: LwQ-Reasoner-10B can retrieve and synthesize information from its trained data to answer factual questions, assist in research, and support knowledge-intensive tasks.

Limitations

Performance Variability Across Languages: While the model supports multiple languages, its performance may vary depending on the language, with better results for languages more prevalent in its training data.
Handling of Niche Topics: The model may struggle to provide accurate information or generate high-quality content for highly specialized or niche topics not covered extensively in its training data.
Complex Multi-Step Reasoning: Although trained on reasoning datasets, the model may still occasionally produce incorrect or incomplete results for multi-step or highly complex reasoning tasks.
Bias and Ethical Concerns: Since LwQ-Reasoner-10B is trained on large, publicly available datasets, it may inherit biases present in the data, leading to potential ethical concerns or inappropriate outputs in certain contexts.
Context Limitations: The model has a finite context window, which may lead to incomplete understanding or response generation for tasks requiring extensive context or very long input texts.
Resource Intensive: As a large-scale model with 10 billion parameters, it requires substantial computational resources for both inference and deployment, limiting its use in resource-constrained environments.
Instruction Ambiguity: The model’s performance can degrade when instructions are ambiguous, vague, or conflicting, potentially leading to outputs that do not align with user expectations.