A Pythia Chat Model of 31M Parameters
Recommended prompt format
<|im_start|>system
{system_message}<|im_end|>
<|im_start|>user
{user_message}<|im_end|>
<|im_start|>assistant
Recommended inference parameters
penalty_alpha: 0.5
top_k: 2
repetition_penalty: 1.0016
Datasets and parameters used for training
SFTTrainer(
model,
train_dataset=train_dataset,
dataset_text_field="text",
eval_dataset=eval_dataset,
max_seq_length=2048,
packing=True,
args=TrainingArguments(
learning_rate=2e-6,
per_device_train_batch_size=1,
per_device_eval_batch_size=1,
gradient_accumulation_steps=16,
lr_scheduler_type="cosine",
num_train_epochs=1,
logging_strategy="steps",
save_strategy="steps",
evaluation_strategy="steps",
logging_steps=10,
eval_steps=10,
save_steps=10,
warmup_steps=50,
load_best_model_at_end=True,
metric_for_best_model="eval_loss",
greater_is_better=False,
weight_decay=0.01,
save_total_limit=10,
neftune_noise_alpha=5,
),
callbacks=[
EarlyStoppingCallback(
early_stopping_patience=3,
early_stopping_threshold=0.005
),
],
)
DPOTrainer(
model,
beta=0.1,
train_dataset=dataset,
tokenizer=tokenizer,
eval_dataset=eval_dataset,
max_length=1536,
max_prompt_length=1024,
args=TrainingArguments(
learning_rate=2e-6,
per_device_train_batch_size=1,
per_device_eval_batch_size=1,
gradient_accumulation_steps=1,
lr_scheduler_type="cosine",
num_train_epochs=1,
logging_strategy="steps",
save_strategy="steps",
evaluation_strategy="steps",
logging_steps=1,
eval_steps=1,
save_steps=1,
warmup_steps=0,
load_best_model_at_end=True,
metric_for_best_model="eval_loss",
greater_is_better=False,
weight_decay=0.0,
neftune_noise_alpha=5,
remove_unused_columns=False,
),
callbacks=[
EarlyStoppingCallback(
early_stopping_patience=3,
early_stopping_threshold=0.005
),
],
)
Detailed results can be found here
| Metric | Value |
|---|
| Avg. | 19.92 |
| AI2 Reasoning Challenge (25-Shot) | 22.70 |
| HellaSwag (10-Shot) | 25.60 |
| MMLU (5-Shot) | 23.24 |
| TruthfulQA (0-shot) | 0.00 |
| Winogrande (5-shot) | 47.99 |
| GSM8k (5-shot) | 0.00 |