SciReason-LFM2-2.6B

SciReason-LFM2-2.6B

<img src="banner.png" />

---

Model Overview

SciReason-LFM2-2.6B is a fine-tuned version of LiquidAI/LFM2-2.6B, trained with Unsloth on the OpenScienceReasoning-2 dataset.
The fine-tuning enhances the base model’s ability to handle multi-step scientific reasoning and produce coherent chain-of-thought explanations.

---

Training Configuration

• Framework: Unsloth
• Dataset: nvidia/OpenScienceReasoning-2
• Examples: ~11,000
• Epochs: 1
• Total Steps: 1,375
• Batch size per device: 2
• Gradient Accumulation Steps: 4
• Effective Batch Size: 8
• Trainable Parameters: ~20M (LoRA / PEFT with Unsloth smart offloading)
• Optimizer: AdamW
• Learning Rate: 2e-4
• Weight Decay: 0.01
• LR Scheduler: cosine with warmup
• Hardware: Single GPU (Unsloth offloading enabled)

---

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model_id = "yasserrmd/SciReason-LFM2-2.6B"
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="bfloat16",
#    attn_implementation="flash_attention_2" <- uncomment on compatible GPU
)
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Generate answer
prompt = """	
Solve the following problem. Make sure to put the answer (and only answer) inside \boxed{}.

Based on analysis of multinational aeromedical databases (e.g., EASA's EMPR, FAA's CAMI database, and military longitudinal studies), which statement accurately characterizes a fundamental limitation in definitively establishing cause-and-effect relationships for cardiovascular morbidity trends among commercial aircrew?

A: Stratified sampling protocols universally eliminate survivorship bias
B: Retroactive harmonization of biochemical markers across jurisdictions enables precise meta-analysis
C: Inability to fully adjust for dominant confounding variables (e.g., socioeconomic status, undisclosed supplement use)
D: Cohort studies consistently show declining age-adjusted myocardial infarction rates compared to the general population
E: Mandatory polysomnography data provides complete correction for sleep disorder comorbidities
F: Radiation dose metrics exhibit a linear correlation with arrhythmia incidence in jet aircraft pilots
G: Genome-wide association studies have identified fully penetrant monogenic risk variants specific to aviators
H: Continuous blood pressure monitoring during all flight phases yields statistically significant longitudinal datasets
I: Pharmacokinetic interactions between hypoxia and statins are conclusively established in CRF models
J: Regulatory divergence causes morbidity rates to universally decline across all regions after 2018"""
input_ids = tokenizer.apply_chat_template(
    [{
    "role":"system",
    "content":""" 
    You are a reasoning assistant.

When solving problems:
- Always place your reasoning inside think tags.
- Think in structured steps, but keep it concise (3–4 short steps maximum).
- Avoid repeating yourself or giving unnecessary background.
- Use bullet points or brief numbered steps for clarity inside think tag.
- After think end tag, provide only the final answer clearly and directly.
- Do not include reasoning outside of the think tags.


    """
},
        {"role": "user", "content": prompt}],
    add_generation_prompt=True,
    return_tensors="pt",
    tokenize=True,
).to(model.device)

output = model.generate(
    input_ids,
    do_sample=True,
    temperature=0.3,
    min_p=0.15,
    repetition_penalty=1.05,
    max_new_tokens=1024,
)

print(tokenizer.decode(output[0], skip_special_tokens=False))

# <|startoftext|><|im_start|>user
# What is C. elegans?<|im_end|>
# <|im_start|>assistant
# C. elegans, also known as Caenorhabditis elegans, is a small, free-living
# nematode worm (roundworm) that belongs to the phylum Nematoda.

---

Intended Use

This model is designed for:

• Scientific reasoning tasks
• Educational Q&A
• Step-by-step logical problem solving

⚠️ Disclaimer: Not intended for clinical or legal decision-making.

---

License

Apache-2.0 License. See LICENSE.

---

Acknowledgements

• LiquidAI for LFM2-2.6B
• NVIDIA for OpenScienceReasoning-2
• Unsloth for efficient fine-tuning with gradient offloading