Zagreus-0.4B-spa

Zagreus-0.4B-spa is a bilingual English/Spanish foundational Small Language Model (SLM) trained from scratch by the mii-llm community (Made in Italy – Large Language Model) on the Seeweb HPC infrastructure.

This is a base (pre-trained) model — it is not instruction-tuned and is intended for researchers, developers, and practitioners who want to fine-tune or build upon a high-quality bilingual English/Spanish foundation.

The Zagreus family represents one of the few openly released, high-performing small language models dedicated to European Romance languages, trained entirely from first principles with a fully transparent pipeline.

---

Model Details

Property	Value
Architecture	Modified Llama-3.2 (fully dense)
Parameters	~400M
Hidden size	960
Intermediate size	2560
Layers	32
Attention heads	15 (KV heads: 5)
Activation	SiLU
Context length	4096 tokens
Tokenizer	Llama-3.2 (vocab_size: 128,256)
Positional encoding	RoPE (theta: 10000.0)
Tied embeddings	Yes
Precision	BF16
Languages	English (400B tokens), Spanish (400B tokens)
Training tokens	~1 trillion
Training framework	Nanotron (mii-llm fork)
Infrastructure	64× NVIDIA A100 GPUs (8 nodes × 8 GPUs), Seeweb HPC

---

Training Data

All datasets used are fully open source and released by Hugging Face:

Dataset	Tokens	Description
FineWeb (350BT sample)	~350B	High-quality English web text
FineWeb-2 (spa_Latn)	—	Spanish web text
FinePDFs (spa_Latn)	—	Spanish PDF documents
StarCoder Data	~250B	Multilingual code

Token distribution: ~400B English + ~400B Spanish + ~200B Code ≈ 1 trillion tokens total

Tokenization

Raw datasets were tokenized using the Llama-3.2 tokenizer (meta-llama/Llama-3.2-1B) via the datatrove library. The process ran for over three weeks of continuous computation on CPU nodes via Slurm, generating approximately 3–5 TB of tokenized data shards.

---

Architecture Choice

We adopted a modified Llama-3.2 fully dense architecture. The choice of a dense model over Mixture-of-Experts (MoE) in the small-parameter regime (~500M) was deliberate: in tightly constrained capacity settings, routing overhead and expert under-utilization typical of MoE architectures may offset their theoretical efficiency advantages. Dense models provide better compute utilization and more stable training dynamics at this scale.

---

Pre-training Configuration

Full Nanotron YAML configuration used for training:

checkpoints:
  checkpoint_interval: 5000
  checkpoints_path: checkpoints_zagreus_spa
  checkpoints_path_is_shared_file_system: false
  save_final_state: false
  save_initial_state: false
data_stages:
- data:
    dataset:
      dataset_folder:
      - /training/pretraining/fineweb-spa/tokenized
      - /training/pretraining/fineweb-edu-350BT/000_tokenized_output
      - /training/pretraining/fineweb-edu-350BT/011_tokenized_output
      - /training/pretraining/fineweb-edu-350BT/012_tokenized_output
      - /training/pretraining/fineweb-edu-350BT/013_tokenized_output
      - /training/pretraining/fineweb-edu-350BT/014_tokenized_output
      - /training/pretraining/fineweb-edu-350BT/015_tokenized_output
      - /training/pretraining/fineweb-edu-350BT/016_tokenized_output
      - /training/pretraining/finepdf-spa/000_tokenized_output
      - /training/pretraining/starcoder_tokenized/000_tokenized_output
    num_loading_workers: 0
    seed: 8
  name: stable phase
  start_training_step: 1
general:
  benchmark_csv_path: null
  consumed_train_samples: null
  ignore_sanity_checks: true
  project: zagreus
  run: zagreus-350M-spa
  seed: 8
  step: null
logging:
  iteration_step_info_interval: 1
  log_level: info
  log_level_replica: info
model:
  ddp_bucket_cap_mb: 100
  dtype: bfloat16
  init_method:
    std: 0.03227
  make_vocab_size_divisible_by: 1
  model_config:
    bos_token_id: 128000
    eos_token_id: 128001
    hidden_act: silu
    hidden_size: 960
    initializer_range: 0.02
    intermediate_size: 2560
    is_llama_config: true
    max_position_embeddings: 4096
    num_attention_heads: 15
    num_hidden_layers: 32
    num_key_value_heads: 5
    pad_token_id: null
    pretraining_tp: 1
    rms_norm_eps: 1.0e-05
    rope_interleaved: false
    rope_scaling: null
    rope_theta: 10000.0
    tie_word_embeddings: true
    use_cache: true
    vocab_size: 128256
optimizer:
  accumulate_grad_in_fp32: true
  clip_grad: 1.0
  learning_rate_scheduler:
    learning_rate: 0.003
    lr_decay_starting_step: 750000
    lr_decay_steps: 50000
    lr_decay_style: linear
    lr_warmup_steps: 4000
    lr_warmup_style: linear
    min_decay_lr: 1.0e-7
  optimizer_factory:
    adam_beta1: 0.9
    adam_beta2: 0.95
    adam_eps: 1.0e-08
    name: adamW
    torch_adam_is_fused: true
  weight_decay: 0.01
  zero_stage: 0
parallelism:
  dp: 64
  expert_parallel_size: 1
  pp: 1
  pp_engine: 1f1b
  recompute_layer: false
  tp: 1
  tp_linear_async_communication: true
  tp_mode: REDUCE_SCATTER
  tp_recompute_allgather: true
profiler: null
tokenizer:
  tokenizer_max_length: null
  tokenizer_name_or_path: meta-llama/Llama-3.2-1B
  tokenizer_revision: null
tokens:
  batch_accumulation_per_replica: 1
  limit_test_batches: 0
  limit_val_batches: 0
  micro_batch_size: 4
  sequence_length: 4096
  train_steps: 2000000
  val_check_interval: 5000

Slurm Launch Script

#SBATCH --job-name=350_es
#SBATCH --account=YOUR_ACCOUNT
#SBATCH --partition=PARTITION
#SBATCH --nodes=8
#SBATCH --gres=gpu:8            # 8 A100 per node = 64 total
#SBATCH --cpus-per-task=32
#SBATCH --time=4-00:00:00
#SBATCH --output=slurm-%j.out

################ 0. Environment ################
module purge
module load profile/global
module load python/3.11 cuda/12.2 cudnn nccl gcc

source /path/to/venv/nanotron/bin/activate

export HF_HOME=/path/to/hf_home
export TRANSFORMERS_OFFLINE=1
export HF_HUB_OFFLINE=1
export HF_DATASETS_OFFLINE=1
export OMP_NUM_THREADS=$SLURM_CPUS_PER_TASK
export NCCL_IB_DISABLE=0
export NCCL_SOCKET_IFNAME="ib0,eno,eth"
export WANDB_MODE=disabled

################ 1. Distributed vars ############
GPUS_PER_NODE=4
NNODES=$SLURM_JOB_NUM_NODES
NODE_RANK=$SLURM_NODEID
MASTER_ADDR=$(scontrol show hostnames $SLURM_JOB_NODELIST | head -n1)
MASTER_PORT=29400
RDZV_ID=$SLURM_JOB_ID

################ 2. Launch ######################
srun torchrun \
      --nnodes $NNODES \
      --nproc_per_node $GPUS_PER_NODE \
      --rdzv_id $RDZV_ID \
      --rdzv_backend c10d \
      --rdzv_endpoint $MASTER_ADDR:$MASTER_PORT \
      /path/to/nanotron/run_train.py \
      --config-file smollm2/zagreus_350M_spa.yaml

Checkpoint Conversion to Hugging Face Format

torchrun --nproc_per_node=1 -m examples.llama.convert_nanotron_to_hf \
  --checkpoint_path=checkpoints/<step> \
  --save_path=hf_checkpoints/<step> \
  --tokenizer_name meta-llama/Llama-3.2-1B

---

Evaluation

Evaluation Commands

lm-eval --model hf --model_args pretrained=<checkpoint> \
  --tasks m_mmlu_es --num_fewshot 5 --device cuda:0 --batch_size 1

lm-eval --model hf --model_args pretrained=<checkpoint> \
  --tasks hellaswag_es,arc_es --device cuda:0 --batch_size 1

Checkpoint Progression

The table below tracks benchmark scores across training checkpoints, demonstrating consistent improvement throughout pre-training across all three Spanish benchmarks:

Checkpoint	MMLU ES ↑	ARC ES ↑	HellaSwag ES ↑	Average
146k	0.254	0.265	0.409	0.309
216k	0.237	0.270	0.414	0.307
292k	0.254	0.262	0.417	0.311
406k	0.254	0.269	0.423	0.315
518k	0.255	0.280	0.429	0.321

Best overall checkpoint: 518k with an average of 0.321 across all three benchmarks. HellaSwag ES shows the clearest and most consistent upward trend throughout training, rising from 0.409 at 146k to 0.429 at 518k (+0.020).

---

Usage

This is a base model — it performs causal language modelling (text completion) and is not instruction-tuned. It is best suited as a starting point for fine-tuning.

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "mii-llm/zagreus-0.4B-spa"

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

# Base model: text completion, not instruction following
prompt = "La inteligencia artificial es una disciplina que"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

output = model.generate(
    **inputs,
    max_new_tokens=200,
    temperature=0.8,
    do_sample=True,
    repetition_penalty=1.1
)

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

---

Full Model Family

Base Models (Zagreus)

Model	Languages	HuggingFace
Zagreus-0.4B-ita	English + Italian	🤗 Link
Zagreus-0.4B-spa (this model)	English + Spanish	🤗 Link
Zagreus-0.4B-por	English + Portuguese	🤗 Link
Zagreus-0.4B-fra	English + French	🤗 Link

Post-trained Models (Nesso) — English/Italian

Model	Use Case	HuggingFace
Nesso-0.4B-instruct	Conversational / Instruction following	🤗 Link
Nesso-0.4B-agentic	Function calling / Agentic	🤗 Link
Open-Zagreus-0.4B	Fully open source	🤗 Link

---

Citation

If you use this model in your research, please cite:

@misc{zagreus2025,
  title        = {The Joy and Pain of Training an LLM from Scratch:
                  A Technical Report on the Zagreus and Nesso Model Families},
  author       = {mii-llm community},
  year         = {2025},
  howpublished = {\url{https://github.com/mii-llm/zagreus-nesso-slm}},
}

---

Acknowledgements

• Antonio Baldassarra (CEO, Seeweb) and Marco Cristofanilli (Head of AI, Seeweb) for commissioning and sponsoring the infrastructure
• The Hugging Face team for Nanotron, datatrove, FineWeb, FineWeb-2, and FinePDFs
• The mii-llm open-source community for contributions to multilingual evaluation harnesses and the Nanotron fork

---

License

Released under the Apache 2.0 license.

Made with ❤️ in Italy by mii-llm