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
chat_template: phi_3

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_nosys_phi
    train_on_split: train

val_set_size: 0.05

# lora
adapter: lora
lora_r: 512
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-nosys-gpt4ominiplans-27k-512rank-long-chattemp

output_dir: ./outputs/phi3-nosys-gpt4ominiplans-27k-512rank-long-chattemp

gradient_accumulation_steps: 2
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: true
micro_batch_size: 2
num_epochs: 3
eval_batch_size: 2
warmup_ratio: 0.05
learning_rate: 2e-5
lr_scheduler: cosine
optimizer: adamw_torch

hub_model_id: verifiers-for-code/phi3-nosys-gpt4ominiplans-27k-512rank-long-chattemp
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-nosys-gpt4ominiplans-27k-512rank-long-chattemp

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: 0.6042

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: 2e-05
• 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: 44
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss
1.0833	0.0034	1	1.0330
0.9533	0.1279	38	0.9330
0.779	0.2559	76	0.7325
0.6894	0.3838	114	0.6651
0.6137	0.5118	152	0.6365
0.6492	0.6397	190	0.6247
0.6231	0.7677	228	0.6190
0.6428	0.8956	266	0.6151
0.6111	1.0236	304	0.6120
0.6138	1.1515	342	0.6101
0.5938	1.2795	380	0.6084
0.5592	1.4074	418	0.6071
0.6041	1.5354	456	0.6062
0.5768	1.6633	494	0.6055
0.6238	1.7912	532	0.6052
0.5944	1.9192	570	0.6048
0.588	2.0471	608	0.6045
0.5989	2.1751	646	0.6045
0.5782	2.3030	684	0.6042
0.5568	2.4310	722	0.6043
0.6032	2.5589	760	0.6043
0.6085	2.6869	798	0.6044
0.5754	2.8148	836	0.6042
0.6106	2.9428	874	0.6042

Framework versions

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