Iker/Llama-3-Instruct-Neurona-8b

Neurona 8B 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 8B 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 8B ha sido entrenado con los siguiente datasets. No en todos los casos se ha usado el dataset completo

• pinzhenchen/alpaca-cleaned-es
• Danielbrdz/Barcenas-Economia
• HiTZ/casimedicos-exp
• somosnlp/coser_resumenes
• csebuetnlp/CrossSum en + es
• Iker/Document-Translation-en-es
• somosnlp/es-inclusive-language-it
• FreedomIntelligence/evol-instruct-spanish
• glaiveai/glaive-code-assistant-v3
• glaiveai/glaive-function-calling-v2
• Iker/InstructTranslation-EN-ES
• somosnlp/lenguaje-claro-dataset
• somosnlp/LingComp_QA
• bltlab/lr-sum
• Iker/NoticIA
• xaviviro/oasst2_es_gpt
• teknium/OpenHermes-2.5
• Iker/OpenHermes-2.5-Spanish
• Helsinki-NLP/opus-100 en es
• projecte-aina/RAG_Multilingual
• sem_eval_2018_task_1
• davidstap/ted_talks
• HiTZ/This-is-not-a-dataset
• wikipedia es

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

Esta es la configuración usada

base_model: meta-llama/Meta-Llama-3-8B-Instruct
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/InstructDatasets/alpaca-cleaned-es.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/Barcenas-Economia.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/casimedicos.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/coser_resumene.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/CrossSum_en.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/CrossSum_es.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/Document-Translation-en-es.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/es-inclusive-language.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/evol-instruct-spanish.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/glaive-code-assistant-v3-small.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/glaive-function-calling-v2.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
        - tool
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/InstructTranslation-EN-ES.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/lenguaje-claro-dataset.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/LingComp_QA.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/lr-sum-es.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/NoticIA.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/NoticIA-large.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/NoticIA-summary.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/oasst2_es_gpt.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/OpenHermes-2.5-English.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/OpenHermes-2.5-Spanish.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/opus-100-en-es.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/RAG_Multilingual-es.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/sem_eval_2018_task_1.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/ted_talks-es_en.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/This-is-not-a-dataset.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human
  - path: /ikerlariak/igarcia945/InstructDatasets/wikipedia-es.jsonl
    type: sharegpt
    conversation: llama3
    field: conversations
    roles:
      input:
        - system
        - gpt
      output:
        - human

chat_template: llama3

dataset_prepared_path: /ikerlariak/igarcia945/Mortadelo-Filemon/Meta-Llama-3-8B-Instruct-Spanish/dataset

shuffle_merged_datasets: true

val_set_size: 0.005

output_dir: /ikerlariak/igarcia945/Mortadelo-Filemon/Meta-Llama-3-8B-Instruct-Spanish

adapter:
lora_model_dir:

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

tokens:
  - "<tool_call>"
  - "<tool_response>"
  - "<tools>"
  - "</tool_call>"
  - "</tool_response>"
  - "</tools>"
  - "<reserved1>"
  - "<reserved2>"

special_tokens:
   pad_token: <|end_of_text|>

neftune_noise_alpha: 5

wandb_project: Mortadelo&Filemon
wandb_entity: igarciaf
wandb_watch:
wandb_name: Meta-Llama-3-8B-Instruct-Spanish
wandb_log_model: 

gradient_accumulation_steps: 32
micro_batch_size: 2
eval_batch_size: 2
num_epochs: 2
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:

seed: 33