A Lightweight Instruction-Tuned LLM by TNSA AI

The NGen-2 Base 124M Instruct model is a compact transformer-based language model designed for efficient inference, rapid prototyping, education, and low-resource environments. It delivers strong instruction-following performance within a small computational footprint and serves as the foundation of the NGen-2 Series.

NGen-2 Base 124M Instruct is part of the NGen-2 model family developed by TNSA AI.
It is a distilled, instruction-optimized model suited for text generation, reasoning tasks, and edge-level deployment.

• Developer: TNSA
• Funding: TNSA Foundation
• Shared by: Thishyaketh Abimalla
• Model Type: Decoder-only Transformer
• Parameter Count: 124M
• Language: English
• Architecture Sources: ARCH-X 7 Pro, TNSA_Standard
• Training Hardware: 8 × A100 80GB
• License: TNSA OpenWeight License

• Transformer Decoder-Only (GPT-style)
• Multi-Head Attention with ARCH-X optimizations
• TNSA_Standard library for training efficiency
• Enhanced normalization for small-parameter stability

• Large curated English corpus
• Instruction dataset built by TNSA
• Safety-aligned prompts
• Low-toxicity text corpora

• Knowledge distillation from larger NGen-2 checkpoints
• Gradient checkpointing for memory efficiency
• TNSA AI curriculum-learning pipeline
• Custom tokenizer: Tokenize2

• Text generation
• Summarization
• Basic reasoning
• Grammar correction
• Education & learning tools
• Lightweight chatbots
• Edge/embedded NLP tasks
• Prototyping new LLM architectures

• Domain-specific assistants
• Customer support bots
• Coding tutors
• Lightweight agent systems
• Content generation pipelines

• High-stakes decision-making
• Medical or legal advisory
• Misinformation generation
• Highly factual retrieval tasks
• Harmful, dangerous, or extremist content

• Lower factual accuracy than large models
• Occasional hallucinations
• Limited multilingual ability
• Restricted context window compared to larger LLMs

The model may:

• Reproduce biases from web-scale training data
• Reflect cultural skew toward English-dominant sources
• Produce stereotypical associations under certain prompts

To mitigate risks:

• Add post-processing filters
• Use retrieval-augmented generation if factual accuracy is needed
• Always test domain-specific fine-tuned versions before production

• Evaluate bias on diverse datasets
• Perform robustness tests before deployment

• Do not use the model for harmful or malicious applications
• Always disclose model limitations to end-users

The code below demonstrates inference with the model:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "TNSA-AI/N-Gen-2"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

text = "Once upon a time"
inputs = tokenizer.encode(text, return_tensors="pt")
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))