ReDiX/SmolLM2-360M-Instruct-ita

See axolotl config

axolotl version: 0.5.0

base_model: HuggingFaceTB/SmolLM2-360M

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: ./dataforge
    type: chat_template

    field_messages: conversations
    message_field_role: from
    message_field_content: value
  
  - path: HuggingFaceTB/smol-smoltalk
    type: chat_template

    field_messages: messages
    message_field_role: role
    message_field_content: content

chat_template: chatml
dataset_prepared_path: last_run_prepared
val_set_size: 0.1
output_dir: ./outputs/smollm360m

sequence_len: 8192
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true

wandb_project: axolotl
wandb_entity:
wandb_watch:
wandb_name: smollm2
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 4
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 1.0e-03

train_on_inputs: false
group_by_length: false
bf16: true
fp16: 
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 5
xformers_attention:
flash_attention: true

warmup_steps: 10
evals_per_epoch: 4
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  pad_token: "<|im_end|>"
  eos_token: "<|im_end|>"

SmolLM2 360M Instruct ITA

This model is a fine-tuned version of HuggingFaceTB/SmolLM2-360M on the smol-smoltalk dataset and on the ReDiX/DataForge. Our datasets is a mixture of open source italian datasets and ReDiX/everyday-conversations-ita It achieves the following results on the evaluation set:

• Loss: 0.8925

Model description

This model is an experiment to test out the ReDiX/everyday-conversations-ita dataset.

Intended uses & limitations

Simple and very basic chat in italian and english

Training and evaluation data

Model	m_mmlu_it	arc_it	hellaswag_it
Qwen2.5-0.5-Instruct	37.05	27.54	35.73
ReDiX/SmolLM2-360M-Instruct-ita	24.94	28.40	35.96

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.001
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Use adamw_bnb_8bit with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.0003	1	1.3366
1.0595	0.2501	774	1.0840
1.0194	0.5002	1548	1.0139
1.0075	0.7504	2322	0.9701
1.0286	1.0005	3096	0.9269
0.7871	1.2506	3870	0.9111
0.7481	1.5007	4644	0.8960
0.7429	1.7508	5418	0.8925

Framework versions

• Transformers 4.46.2
• Pytorch 2.5.1+cu124
• Datasets 3.1.0
• Tokenizers 0.20.3