Model Description:
Nexa-Qwen-sci-7B is a fine-tuned variant of the open-weight Qwen/Qwen3-1.7B model, optimized for scientific research generation tasks such as hypothesis generation, abstract writing, and methodology completion. Fine-tuning was performed using the PEFT (Parameter-Efficient Fine-Tuning) library with LoRA in 4-bit quantized mode using the bitsandbytes backend. The model leverages Qwen3’s thinking mode (enable_thinking=True) for enhanced reasoning capabilities, making it suitable for complex scientific tasks.

This model is part of the Nexa Scientific Intelligence (Psi) series, developed for scalable, automated scientific reasoning and domain-specific text generation.

Developed by: Allan (Independent Scientific Intelligence Architect)
Funded by: Self-funded
Shared by: Allan (https://huggingface.co/Allanatrix)
Model type: Decoder-only transformer (causal language model)
Language(s): English (scientific domain-specific vocabulary)
License: Apache 2.0 (inherits from base model)
Fine-tuned from: Qwen/Qwen3-1.7B
Repository: https://huggingface.co/allan-wandia/nexa-qwen-sci-7b
Demo: Coming soon via Hugging Face Spaces or Lambda inference endpoint.

Uses

Direct Use

Scientific hypothesis generation
Abstract and method section synthesis
Domain-specific research writing
Semantic completion of structured research prompts

Downstream Use

Fine-tuning or distillation into smaller expert models
Foundation for test-time reasoning agents
Seed model for bootstrapping larger synthetic scientific corpora

Out-of-Scope Use

General conversation or chat use cases
Non-English scientific domains
Legal, financial, or clinical advice generation

Bias, Risks, and Limitations

While the model performs well on structured scientific input, it inherits biases from its base model (Qwen3-1.7B) and fine-tuning dataset. Results should be evaluated by domain experts before use in high-stakes settings. It may hallucinate plausible but incorrect facts, especially in low-data areas. The thinking mode may increase latency for simpler tasks but improves reasoning quality.

Recommendations

Users should:

Validate critical outputs against trusted scientific literature
Avoid deploying in clinical or regulatory environments without further evaluation
Consider additional domain fine-tuning for niche fields
Use recommended sampling parameters (Temperature=0.6, TopP=0.95, TopK=20, MinP=0, Presence Penalty=1.5) to avoid endless repetitions in thinking mode

How to Get Started with the Model

python

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "allan-wandia/nexa-qwen-sci-7b"

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

prompt = "Generate a novel hypothesis in quantum materials research:"
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=32768,
    temperature=0.6,
    top_p=0.95,
    top_k=20,
    min_p=0,
    presence_penalty=1.5
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Training Data Size: 100 million tokens sampled from a 500M+ token corpus Source: Curated scientific literature, abstracts, methodologies, and domain-labeled corpora (Bio, Physics, QST, Astro Labeling: Token-level labels auto-generated via Qwen3 tokenizer with chat template (enable_thinking=True)

Preprocessing

Tokenization with sequence truncation to 32,768 tokens (Qwen3’s context length) Formatted using Qwen3’s chat template with thinking mode enabled Labeled and batched using CPU; inference dispatched to GPU asynchronously Training Hyperparameters Base model: Qwen/Qwen3-1.7B Sequence length: 32768 Batch size: 1 (with gradient accumulation) Gradient Accumulation Steps: 64 Effective Batch Size: 64 Learning rate: 2e-5 Epochs: 2 LoRA: Enabled (PEFT with RSLoRA) Quantization: 4-bit via bitsandbytes Optimizer: 8-bit AdamW Framework: Transformers (≥4.51.0) + PEFT + Accelerate + TRL

Sampling Parameters: Temperature=0.6, TopP=0.95, TopK=20, MinP=0, Presence Penalty=1.5 (applied during inference) Evaluation Testing Data

Synthetic scientific prompts across domains (Physics, biology, and Materials Science)

Evaluation Factors Hypothesis novelty (entropy score) Internal scientific consistency (domain-specific rubric) Reasoning quality (assessed via thinking mode outputs)

Results The model performs robustly in hypothesis generation and scientific prose tasks, with enhanced reasoning capabilities due to Qwen3’s thinking mode. Coherence is high, and novelty depends on prompt diversity. It is well-suited as a distiller or inference agent for synthetic scientific corpora generation.

Environmental Impact

Component Value Hardware Type: 2× NVIDIA T4 GPUs Hours used: ~7.5 Cloud Provider Kaggle (Google Cloud) Compute Region US Carbon Emitted Estimate pending (likely 1 kg COkg CO2)

Technical Specifications

Model Architecture Transformer decoder (Qwen3-1.7B architecture: 28 layers, 16 attention heads for Q, 8 for KV) LoRA adapters applied to all linear layers with RSLoRA

Quantized with bytes to 4-bit for memory efficiency

Compute Infrastructure CPU: Intel i5 8th Gen vPro (batch preprocessing) GPU: 2× NVIDIA T4 (CUDA 12.1) Software Stack PEFT 0.12.0 Transformers 4.51.0 Accelerate TRL Torch 2.x

Citation BibTeX:

@misc{nexa-qwen-sci-7b, title = {Nexa Qwen Sci 7B}, author = {Allan Wandia}, year = {2025}, howpublished = {\url{https://huggingface.co/allan-wandia/nexa-qwen-sci-7b}}, note = {Fine-tuned model for scientific generation tasks with Qwen3 thinking mode} }

Model Card Contact

For questions, contact Allan via Hugging Face or at Email: allanw.mk@gmail.com

Model Card Authors

Allan Wandia (Independent ML Engineer and Systems Architect)

Glossary

LoRA: Low-Rank Adaptation PEFT: Parameter-Efficient Fine-Tuning Entropy Score: Metric used to estimate novelty/variation Safe Tensors: Secure, fast format for model weights Thinking Mode: Qwen3’s feature for enhanced reasoning, enabled via enable_thinking=True

Links Github Repo and notebook: https://github.com/DarkStarStrix/Nexa_Auto

Allanatrix/Nexa-Qwen-sci-7B

Model Card for Nexa-Qwen-sci-7B

Model Details