toroe/SmolLM-3B-Science-EN

# SmolLM3-3B — English Reasoning Instruction Fine-Tune (Nemotron Multilingual Reasoning)

## Model Description

This model is a Supervised Fine-Tuned (SFT) version of:

HuggingFaceTB/SmolLM3-3B

It was trained on the English (`en`) split of:

DGurgurov/Nemotron-Multilingual-Reasoning

The purpose of this fine-tune is to improve:

• —English instruction following
• —multi-step reasoning
• —long-context chat behavior

The dataset was converted into structured chat conversations and optimized using completion-only loss, meaning only the assistant’s responses contributed to the training objective.

### Key Characteristics

• —Base model: SmolLM3-3B
• —Language: English specialization
• —Context length during training: 16,384 tokens
• —Chat formatted conversations
• —Packed sequences
• —Long-context reasoning tuning

## Intended Uses

### Suitable

• —Conversational assistants
• —Instruction-following agents
• —Reasoning tasks
• —Educational tutoring
• —Long-document Q&A
• —Research on small long-context LLMs

### Not Suitable

• —Medical or legal advice
• —Autonomous decision making
• —Safety-critical systems
• —Financial decision automation

## Training Data

Dataset:

DGurgurov/Nemotron-Multilingual-Reasoning

Processing configuration:

• —Language filter: English only
• —Converted to chat messages (prepare_messages=True)
• —Assistant-only loss masking (completion_only_loss=True)

User and system prompts were masked during training; only assistant tokens produced gradients.

Please consult the dataset card for data provenance and limitations.

## Training Procedure

Training used HuggingFace Accelerate with Fully Sharded Data Parallel (FSDP) across 8 processes.

### Core Setup

• —Method: Supervised fine-tuning (SFT)
• —Epochs: 3
• —Max sequence length: 16,384
• —Packing: enabled
• —Precision: bfloat16

• —Gradient checkpointing: enabled
• —Liger kernel: enabled
• —Distributed training: FSDP

### Optimization

• —Optimizer: adamw_torch_fused
• —Batch size per device: 4
• —Gradient accumulation: 4
• —Effective batch size per GPU: 16 sequences / step
• —Weight decay: 0.05

Learning rate schedule:

• —Scheduler: cosine_with_min_lr
• —Warmup ratio: 0.05
• —Minimum learning rate: 5e-6

### Logging & Checkpoints

• —Logging: every 5 steps
• —Checkpoint: every 450 steps
• —Tracking: Weights & Biases
• —Token accuracy logged during training

### Data Processing

• —Dataset preprocessing workers: 16
• —Chat formatting: enabled
• —Dataset preparation: enabled
• —Language split: en

## Usage

### Transformers Example

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

 model_id = "YOUR_USERNAME/YOUR_MODEL_REPO"

 tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

 messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Explain why the sky is blue."}
]

 prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

 outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
)

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

Important: Use apply_chat_template() when prompting. The model was trained on chat-formatted conversations and performance will degrade without it.

## Evaluation

During training, token accuracy was logged as a diagnostic metric.

Token accuracy:

• —helps monitor training stability
• —is not a benchmark score
• —does not measure reasoning quality

For meaningful evaluation, use:

• —instruction-following benchmarks
• —reasoning datasets
• —long-context tasks

## Limitations

• —May hallucinate incorrect information
• —Reasoning chains may contain logical mistakes
• —Performance near 16k tokens depends heavily on prompt structure
• —Smaller model → less world knowledge than large LLMs
• —Not suitable for safety-critical deployment

## Bias & Safety

The model inherits biases from:

• —the base model
• —the training dataset

Recommended mitigations:

• —moderation filtering
• —safety-oriented system prompts
• —human oversight in sensitive use cases

## License

This is a derivative model of:

HuggingFaceTB/SmolLM3-3B

The original base model license and restrictions apply, along with dataset terms.

Verify compatibility before commercial usage.

## Reproducibility (Training Arguments)

accelerate launch --use_fsdp --num_processes 8 --config_file sft/my_config.yaml sft/sft_trainer.py

 --model_name HuggingFaceTB/SmolLM3-3B
--tokenizer_name HuggingFaceTB/SmolLM3-3B
--dataset_path DGurgurov/Nemotron-Multilingual-Reasoning
--skip_prepare_dataset False
--lang_split en
--prepare_messages True
--completion_only_loss True
--max_length 16384

## Citation

If you use this model, please cite:

• —HuggingFaceTB/SmolLM3-3B
• —DGurgurov/Nemotron-Multilingual-Reasoning

## Acknowledgements

• —HuggingFaceTB — SmolLM3 base model
• —Nemotron Multilingual Reasoning dataset authors
• —HuggingFace Accelerate and Transformers libraries