license: mit
base_model: DeepMount00/Mistral-Ita-7b
tags:
- axolotl
- generated_from_trainer
- psycology
- companion
model-index:
- name: Samantha-ita-v0.1
results: []
datasets:
- WasamiKirua/samantha-ita
- WasamiKirua/psycology-dataset-ita
language:
- it
See axolotl config
axolotl version: 0.4.0
base_model: DeepMount00/Mistral-Ita-7b
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
load_in_8bit: false
load_in_4bit: false
strict: false
datasets:
- path: /workspace/datasets/samantha-ita-sharegpt.jsonl
type: sharegpt
field: conversations
- path: /workspace/datasets/psycology-dataset-gpt-ita.jsonl
type: sharegpt
field: conversations
chat_template: chatml
hub_model_id: Samantha-ita-v0.1
dataset_prepared_path:
val_set_size: 0.05
output_dir: ./out
sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true
eval_sample_packing: false
wandb_project: samantha-mistral7b
wandb_entity:
wandb_watch:
wandb_name: Samantha-ita-v0.1
wandb_log_model:
gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.000006
# 0.000006 OK better curve
# 0.0005 OK
train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false
gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
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:
bos_token: "<s>"
eos_token: "<|im_end|>"
unk_token: "<unk>"
tokens:
- "<|im_start|>"
- "<|im_end|>"
Samantha-ita-v0.1
This model is a fine-tuned version of DeepMount00/Mistral-Ita-7b on the None dataset. It achieves the following results on the evaluation set:
- Loss: 0.7069
Model description
Samantha is a fine-tuned Italian version based on Eric Hartford's Samantha. For this, I utilized the pre-trained Mistral 7B version. The model performs excellently! Please take a look at the datasets used.
Intended uses & limitations
Sure, here's the corrected and improved version:
Samantha is a proficient companion who understands and speaks Italian fluently. She has undergone training on various topics. In addition to the original Samantha dataset translated with GPT-4, I have also incorporated a psychology conversations dataset to further enrich Samantha's knowledge in the field of psychology."
Chat Template
<|im_start|>system
YOUR PROMPT<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant
Quantized Versions:
GGUF availabile here: https://huggingface.co/WasamiKirua/Samantha-ita-mistral-v0.1-GGUF
DPO Version
DPO trained version available here: https://huggingface.co/WasamiKirua/Samantha-ita-mistral-v0.1-DPO
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 6e-06
- train_batch_size: 2
- eval_batch_size: 2
- seed: 42
- gradient_accumulation_steps: 4
- total_train_batch_size: 8
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 10
- num_epochs: 2
Training results
| Training Loss | Epoch | Step | Validation Loss |
|---|---|---|---|
| 1.9261 | 0.01 | 1 | 1.8998 |
| 0.8902 | 0.25 | 28 | 0.8267 |
| 0.8422 | 0.5 | 56 | 0.7604 |
| 0.8338 | 0.75 | 84 | 0.7299 |
| 0.8397 | 1.0 | 112 | 0.7136 |
| 0.6859 | 1.22 | 140 | 0.7131 |
| 0.6707 | 1.47 | 168 | 0.7082 |
| 0.7041 | 1.72 | 196 | 0.7069 |
| 0.6936 | 1.97 | 224 | 0.7069 |
Framework versions
- Transformers 4.40.0.dev0
- Pytorch 2.2.0
- Datasets 2.15.0
- Tokenizers 0.15.0