lina5555/unmaskednemo

See axolotl config

axolotl version: 0.8.0

base_model: mistralai/Mistral-Nemo-Instruct-2407
model_type: MistralForCausalLM
hub_model_id: Alignment-Lab-AI/linabot
strict: false
chat_template: tokenizer_default
plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_layer_norm: true
liger_fused_linear_cross_entropy: true
datasets:
  - path: linabot/train_data
    type: chat_template
    field_messages: messages
    message_property_mappings:
      role: role
      content: content
    roles_to_train: ['assistant', 'user']
    train_on_eos: turn

learning_rate: 2e-5
lr_scheduler: cosine
weight_decay: 0.03
warmup_steps: 450
dataset_prepared_path:
val_set_size: 0.2
output_dir: ./outputs/out

sequence_len: 10400
sample_packing: true
pad_to_sequence_len: true
eval_sample_packing: true

wandb_project: linabot
wandb_entity:
wandb_watch: all
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 5
optimizer: adalomo
lr_scheduler: cosine
learning_rate: 0.0002024
flash_attention: true
flash_attn_cross_entropy: false
flash_attn_rms_norm: true
flash_attn_fuse_qkv: false
flash_attn_fuse_mlp: true
torch_compile_mode: "max-autotune"
bf16: auto
tf32: false

gradient_checkpointing: true
resume_from_checkpoint:
logging_steps: 1

evals_per_epoch: 8
saves_per_epoch: 1
weight_decay: 0.03
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  pad_token: "<pad>"

linabot

This model is a fine-tuned version of mistralai/Mistral-Nemo-Instruct-2407 on the linabot/train_data dataset. It achieves the following results on the evaluation set:

• Loss: 0.0558

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002024
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• optimizer: Use OptimizerNames.ADALOMO and the args are: No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 450
• num_epochs: 5.0

Training results

Training Loss	Epoch	Step	Validation Loss
1.526	0.0083	1	1.5474
1.5934	0.125	15	1.5472
1.5242	0.25	30	1.5454
1.5296	0.375	45	1.5408
1.5087	0.5	60	1.5322
1.486	0.625	75	1.5188
1.4314	0.75	90	1.5005
1.4311	0.875	105	1.4782
1.4532	1.0	120	1.4513
1.4215	1.125	135	1.4198
1.3248	1.25	150	1.3825
1.2697	1.375	165	1.3386
1.3281	1.5	180	1.2880
1.2428	1.625	195	1.2296
1.1533	1.75	210	1.1596
1.1038	1.875	225	1.0747
1.0226	2.0	240	0.9723
0.8858	2.125	255	0.8467
0.6762	2.25	270	0.7047
0.6433	2.375	285	0.5626
0.4017	2.5	300	0.4283
0.2875	2.625	315	0.3072
0.2244	2.75	330	0.2161
0.1445	2.875	345	0.1572
0.0898	3.0	360	0.1192
0.0666	3.125	375	0.0991
0.0605	3.25	390	0.0855
0.0457	3.375	405	0.0757
0.052	3.5	420	0.0700
0.0634	3.625	435	0.0658
0.0364	3.75	450	0.0623
0.045	3.875	465	0.0601
0.0395	4.0	480	0.0582
0.0558	4.125	495	0.0573
0.0468	4.25	510	0.0566
0.0399	4.375	525	0.0562
0.0337	4.5	540	0.0560
0.0413	4.625	555	0.0559
0.0318	4.75	570	0.0558
0.0435	4.875	585	0.0558
0.0445	5.0	600	0.0558

Framework versions

• Transformers 4.51.1
• Pytorch 2.6.0+cu124
• Datasets 3.5.0
• Tokenizers 0.21.1