Neurona-2b

Neurona 2B Beta: Un Modelo de Lenguage en Español

Esta es una versión preliminar del dataset card. El modelo está en desarrollo y no es la versión final. Si quieres saber más sobre este modelo, escribe a [email protected]

Neurona 2B es un modelo de lenguaje en Español. Esta es la primera iteración y un experimento para poner a punto los scripts y la infraestructura.

Neurona 2B ha sido entrenado con los siguiente datasets

• teknium/OpenHermes-2.5
• Iker/OpenHermes-2.5-Spanish
• Iker/Document-Translation-en-es
• Iker/InstructTranslation-EN-ES
• Helsinki-NLP/opus-100 (en-es, only a few examples to reach 1 million instructions)
• projecte-aina/RAG_Multilingual(es only, 3701 examples)
• glaiveai/glaive-code-assistant-v3
• glaiveai/glaive-function-calling-v2

Esta mezcla de datasets en Inglés y Español, permite al modelo adquirir diferentes capacidades, como RAG, function calling, code assistant, question answering, summarization... tanto en Inglés como en Español.

Entrenamiento

Este modelo se ha entrado usando 4xNvidia A100 80Gb y axolotl <img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

Esta es la configuración usada

base_model: google/gemma-2b
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
is_falcon_derived_model:
is_llama_derived_model:
is_qwen_derived_model:
is_mistral_derived_model:

load_in_8bit: false
load_in_4bit: false
strict: false

device_map: null

datasets:
  - path: /ikerlariak/igarcia945/Mortadelo-Filemon/final_dataset/OpenHermes-2.5-Spanish_fix_gpt.jsonl
    type: sharegpt
    conversation: chatml
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/Mortadelo-Filemon/final_dataset/OpenHermes-2.5-English.jsonl
    type: sharegpt
    conversation: chatml
    field: conversations
  - path: /ikerlariak/igarcia945/Mortadelo-Filemon/final_dataset/glaive-function-calling-v2.jsonl
    type: sharegpt
    conversation: chatml
    field: conversations
    roles:
      input:
        - system
        - gpt
        - tool
      output:
        - human
  - path: /ikerlariak/igarcia945/Mortadelo-Filemon/final_dataset/glaive-code-assistant-v3-small.jsonl
    type: sharegpt
    conversation: chatml
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
chat_template: chatml

dataset_prepared_path: /ikerlariak/igarcia945/Mortadelo-Filemon/gemma-2b-spanish/dataset

shuffle_merged_datasets: true

val_set_size: 0.005

output_dir: /ikerlariak/igarcia945/Mortadelo-Filemon/gemma-2b-spanish/

adapter:
lora_model_dir:

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

special_tokens:
  bos_token: "<|im_start|>"
  eos_token: "<|im_end|>"
  pad_token: "<|end_of_text|>"

tokens:
  - "<|begin_of_text|>"
  - "<|end_of_text|>"
  - "<|im_start|>"
  - "<|im_end|>"
  - "<|start_header_id|>"
  - "<|end_header_id|>"
  - "<tool_call>"
  - "<tool_response>"
  - "<tools>"
  - "</tool_call>"
  - "</tool_response>"
  - "</tools>"
  - "<reserved1>"
  - "<reserved2>"
  - "<reserved3>"
  - "<reserved4>"



neftune_noise_alpha: 5

wandb_project: Mortadelo&Filemon
wandb_entity: igarciaf
wandb_watch:
wandb_name: gemma2b
wandb_log_model: 

gradient_accumulation_steps: 32
micro_batch_size: 2
eval_batch_size: 2
num_epochs: 3
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 0.00007


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

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_ratio: 0.03
evals_per_epoch: 4
eval_table_size:
save_strategy: "no"
debug:
deepspeed: /ikerlariak/igarcia945/Mortadelo-Filemon/train_configs/deepspeed_zero3.json
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:

seed: 33