# train w/ shisa-ai/shisa-v1-athenev2-reannotated-filtered

base_model: meta-llama/Meta-Llama-3.1-8B-Instruct
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

# User Liger
plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: true

# https://axolotl-ai-cloud.github.io/axolotl/docs/config.html - shards should let us uses half?
# https://github.com/axolotl-ai-cloud/axolotl/discussions/2280
chat_template: llama3
datasets:
  - path: shisa-ai/shisa-v1-athenev2-reannotated-filtered
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    split: train[:50%]
  - path: shisa-ai/shisa-v2-code-math-reasoning-sft-mix
    type: chat_template
    field_messages: conversations
    message_property_mappings:
      role: role
      content: content
    roles:
      system:
        - system
      assistant:
        - gpt
        - model
        - assistant
      user:
        - human
        - user
    roles_to_train: ["system", "input", "assistant"]
    split: train[:50%]
dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: ./outputs/ablation-38-rafathenev2.cmrmix.fixedlength-shisa-v2-llama-3.1-8b-lr8e6

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

# marginal difference
neftune_noise_alpha: 5

use_wandb: true
wandb_project: shisa-v2
wandb_entity: augmxnt
wandb_name: ablation-38-rafathenev2.cmrmix.fixedlength-shisa-v2-llama-3.1-8b-lr8e6

gradient_accumulation_steps: 2
micro_batch_size: 4
num_epochs: 3
optimizer: paged_adamw_8bit
lr_scheduler: linear
learning_rate: 8e-6

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 100
evals_per_epoch: 2
eval_table_size:
saves_per_epoch: 0
save_total_limit: 1 # Only store a single checkpoint
debug:
deepspeed: zero3_bf16.json
weight_decay: 0.00
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|end_of_text|>

outputs/ablation-38-rafathenev2.cmrmix.fixedlength-shisa-v2-llama-3.1-8b-lr8e6

This model is a fine-tuned version of meta-llama/Meta-Llama-3.1-8B-Instruct on the shisa-ai/shisa-v1-athenev2-reannotated-filtered and the shisa-ai/shisa-v2-code-math-reasoning-sft-mix datasets. It achieves the following results on the evaluation set:

Loss: 0.6626

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: 8e-06
train_batch_size: 4
eval_batch_size: 4
seed: 42
distributed_type: multi-GPU
num_devices: 8
gradient_accumulation_steps: 2
total_train_batch_size: 64
total_eval_batch_size: 32
optimizer: Use OptimizerNames.PAGED_ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 100
num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss
0.8916	0.0056	1	0.8836
0.6897	0.5028	89	0.7048
0.6052	1.0056	178	0.6718
0.6128	1.5085	267	0.6615
0.5137	2.0113	356	0.6573
0.514	2.5141	445	0.6626

Framework versions

Transformers 4.48.3
Pytorch 2.6.0+cu124
Datasets 3.2.0
Tokenizers 0.21.0

shisa-ai/ablation-38-rafathenev2.cmrmix.fixedlength-shisa-v2-llama-3.1-8b-lr8e6

outputs/ablation-38-rafathenev2.cmrmix.fixedlength-shisa-v2-llama-3.1-8b-lr8e6

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

shisa-ai/
ablation-38-rafathenev2.cmrmix.fixedlength-shisa-v2-llama-3.1-8b-lr8e6