_ (`-.  ('-. .-.   ('-.      .-')   _   .-')            _ .-') _  
  ( (OO  )( OO )  /  ( OO ).-. ( OO ).( '.( OO )_         ( (  OO) )  
 _.`     \,--. ,--.  / . --. /(_)---\_),--.   ,--.) ,-.-') \     .'_  
(__...--''|  | |  |  | \-.  \ /    _ | |   `.'   |  |  |OO),`'--..._)
 |  /  | ||   .|  |.-'-'  |  |\  :` `. |         |  |  |  \|  |  \  '  
 |  |_.' ||       | \| |_.'  | '..`''.)|  |'.'|  |  |  |(_/|  |   ' |  
 |  .___.'|  .-.  |  |  .-.  |.-._)   \|  |   |  | ,|  |_.'|  |   / :  
 |  |     |  | |  |  |  | |  |\       /|  |   |  |(_|  |   |  '--'  / 
 `--'     `--' `--'  `--' `--' `-----' `--'   `--'  `--'   `-------'  

See axolotl config

axolotl version: 0.3.0

base_model: microsoft/phi-2
model_type: PhiForCausalLM
tokenizer_type: AutoTokenizer
is_llama_derived_model: false
trust_remote_code: true

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: SE6446/SE6446_phasmid_ds
    type: completion

hub_model_id: SE6446/Phasmid-2_v2
hub_strategy: every_save
use_auth_token: true
dataset_prepared_path: /phasmid-2-ds-path
val_set_size: 0.05
output_dir: ./phasmid-sft-out

sequence_len: 2048
sample_packing: true
pad_to_sequence_len:

adapter:
lora_model_dir:
lora_r:
lora_alpha:
lora_dropout:
lora_target_linear:
lora_fan_in_fan_out:

wandb_project: 
wandb_entity: 
wandb_watch:
wandb_name: 
wandb_log_model:

gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 4
optimizer: adamw_torch
adam_beta2: 0.95
adam_epsilon: 0.00001
max_grad_norm: 1.0
lr_scheduler: cosine
learning_rate: 0.0003

train_on_inputs: false
group_by_length: true
bf16: true
fp16: false
tf32: true

gradient_checkpointing:
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention:

warmup_steps: 100
evals_per_epoch: 4
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.1
fsdp:
fsdp_config:
resize_token_embeddings_to_32x: true
special_tokens:
  bos_token: "<|endoftext|>"
  eos_token: "<|endoftext|>"
  unk_token: "<|endoftext|>"
  pad_token: "<|endoftext|>"

Phasmid-2_v2

This model is a fine-tuned version of microsoft/phi-2 on a mix of no_robots and the PIPPA dataset. It achieves the following results on the evaluation set:

• Loss: 2.2924

Model description

Phasmid-2 has been trained on intructional data and thus can perform far better at instruction following than phi-2. However I have not extensively tested the model.

Intended uses & limitations

This model is little more than a side project and I shall treat it as such. Phasmid-2 (due to it's size), can still suffer from problematic hallucinations and poor information. No effort was made to reduce potentially toxic responses, as such you should train this model further if you require it to do so.

Inference

Ensure that eniops is installed

pip install einops

Phi doesn't like device_map = auto, therefore you should specify as like the following:

1. FP16 / Flash-Attention / CUDA:

   model = AutoModelForCausalLM.from_pretrained("SE6446/Phasmid-2_v2", torch_dtype="auto", flash_attn=True, flash_rotary=True, fused_dense=True, device_map="cuda", trust_remote_code=True)
   

2. FP16 / CUDA:

   model = AutoModelForCausalLM.from_pretrained("SE6446/Phasmid-2_v2", torch_dtype="auto", device_map="cuda", trust_remote_code=True)
   

3. FP32 / CUDA:

   model = AutoModelForCausalLM.from_pretrained("SE6446/Phasmid-2_v2", torch_dtype=torch.float32, device_map="cuda", trust_remote_code=True)
   

4. FP32 / CPU:

   model = AutoModelForCausalLM.from_pretrained("SE6446/Phasmid-2_v2", torch_dtype=torch.float32, device_map="cpu", trust_remote_code=True)
   

And then use the following snippet

tokenizer = AutoTokenizer.from_pretrained("SE6446/Phasmid-2_v2", trust_remote_code=True, torch_dtype="auto")
inputs = tokenizer('''SYSTEM: You are a helpful assistant. Please answer truthfully and politely. {custom_prompt}\n
                      USER: {{userinput}}\n
                      ASSISTANT: {{character name if applicable}}:''', return_tensors="pt", return_attention_mask=False)
outputs = model.generate(**inputs, max_length=200)
text = tokenizer.batch_decode(outputs)[0]
print(text)

it should generate after "ASSISTANT:".

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• optimizer: Adam with betas=(0.9,0.95) and epsilon=1e-05
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss
2.3313	0.0	1	2.1374
2.5755	0.25	1319	2.5281
2.4864	0.5	2638	2.5314
2.0961	0.75	3957	2.4697
2.6547	1.0	5276	2.4213
2.1235	1.24	6595	2.3926
1.8875	1.49	7914	2.3233
0.9059	1.74	9233	2.2590
2.2046	1.99	10552	2.1985
1.1938	2.23	11871	2.2555
1.1425	2.48	13190	2.2393
0.6688	2.73	14509	2.2237
1.1111	2.98	15828	2.2126
0.651	3.21	17147	2.2859
0.8669	3.46	18466	2.2914
0.4149	3.71	19785	2.2924

Framework versions

• Transformers 4.37.0.dev0
• Pytorch 2.0.1+cu118
• Datasets 2.16.1
• Tokenizers 0.15.0