See axolotl config

axolotl version: 0.4.1

# model and tokenizer
base_model: microsoft/Phi-3-mini-4k-instruct # change for model
trust_remote_code: true
sequence_len: 2048

strict: false

model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
bf16: auto
pad_to_sequence_len: true
save_safetensors: true


datasets:
  - path: verifiers-for-code/sampled_10k_from_27k
    type: completion
    field: text
    train_on_split: train

val_set_size: 0.05

# lora
adapter: lora
lora_r: 256
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true
lora_modules_to_save:
  - embed_tokens
  - lm_head
use_rslora: true

# logging
wandb_project: valeris
wandb_name: phi3-gpt4ominiplans-27k-256rank

output_dir: ./outputs/phi3-gpt4ominiplans-27k-256-rank

gradient_accumulation_steps: 2
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: true
micro_batch_size: 2
num_epochs: 1
eval_batch_size: 2
warmup_ratio: 0.05
learning_rate: 5e-6
lr_scheduler: cosine
optimizer: adamw_torch

hub_model_id: verifiers-for-code/phi3-gpt4ominiplans-27k-256rank
push_to_hub: true
hub_strategy: all_checkpoints
hub_always_push: true
evals_per_epoch: 8
saves_per_epoch: 4
logging_steps: 1
# eval_table_size: 10
# eval_max_new_tokens: 512

tokens: ["<thinking>", "</thinking>", "<plan>", "</plan>"]

special_tokens:
  pad_token: "<|endoftext|>"

phi3-gpt4ominiplans-27k-256rank

This model is a fine-tuned version of microsoft/Phi-3-mini-4k-instruct on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.8897

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: 5e-06
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• total_eval_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 14
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
1.3503	0.0034	1	1.3120
1.2397	0.1279	38	1.2282
1.1796	0.2559	76	1.1337
1.0737	0.3838	114	1.0539
0.9599	0.5118	152	0.9841
0.9443	0.6397	190	0.9240
0.8986	0.7677	228	0.8931
0.9255	0.8956	266	0.8897

Framework versions

• PEFT 0.11.1
• Transformers 4.44.0.dev0
• Pytorch 2.4.0
• Datasets 2.19.1
• Tokenizers 0.19.1