toroe/SmolLM-3B-Science-ES

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

Model Description

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

HuggingFaceTB/SmolLM3-3B

Fine-tuned on the Spanish (es) split of:

DGurgurov/Nemotron-Multilingual-Reasoning

The goal of this training run was to improve:

• Spanish instruction following
• multi-step reasoning
• conversational behavior
• long-context understanding

Training used structured chat conversations and completion-only loss, meaning only the assistant responses were optimized.

Key Characteristics

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

---

Intended Uses

Suitable

• Spanish conversational assistants
• tutoring or educational assistants
• reasoning and explanation tasks
• document question answering
• research on efficient small LLMs

Not Suitable

• legal or medical advice
• autonomous decision making
• safety-critical systems
• high-risk financial use

---

Training Data

Dataset:

DGurgurov/Nemotron-Multilingual-Reasoning

Processing configuration:

• Language filter: Spanish only
• Converted to chat messages (prepare_messages=True)
• Assistant-only optimization (completion_only_loss=True)

User and system messages were masked during training.

Consult the dataset card for data sources and limitations.

---

Training Procedure

Training was performed using HuggingFace Accelerate with Fully Sharded Data Parallel (FSDP) across 8 processes.

Core Setup

• Method: Supervised fine-tuning (SFT)
• Epochs: 3
• Maximum sequence length: 16,384 tokens
• Sequence 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 steps: 4
• Effective batch size per GPU: 16 sequences per step
• Weight decay: 0.05

Learning rate schedule:

• Scheduler: cosine_with_min_lr
• Warmup ratio: 0.05
• Minimum LR: 5e-6

---

Logging & Checkpoints

• Logging every 5 steps
• Checkpoint every 450 steps
• Weights & Biases tracking
• Token accuracy logged during training

---

Data Processing

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

---

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": "Eres un asistente útil."},
    {"role": "user", "content": "¿Por qué el cielo es azul?"}
]

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:

• monitors training stability
• is not a benchmark
• does not measure reasoning ability

For meaningful evaluation, use:

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

---

Limitations

• May hallucinate incorrect information
• Reasoning chains may contain logical errors
• Performance near 16k tokens depends heavily on prompt structure
• Smaller model → weaker world knowledge than larger 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 review for sensitive applications

---

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 use.

---

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 es
--prepare_messages True
--completion_only_loss True
--max_length 16384
--dataset_num_proc 16
--packing True
--use_liger_kernel True
--bf16 True
--log_token_accuracy True
--optim adamw_torch_fused
--gradient_checkpointing True
--per_device_train_batch_size 4
--gradient_accumulation_steps 4
--ddp_find_unused_parameters False
--lr_scheduler_type cosine_with_min_lr
--lr_scheduler_kwargs {"min_lr": 5.0e-6}
--warmup_ratio 0.05
--weight_decay 0.05
--report_to wandb
--run_name smol_3b_3epochs_lns_es
--num_train_epochs 3
--save_strategy steps
--logging_steps 5
--save_steps 450

---

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