File size: 8,815 Bytes
07423df |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 |
import hashlib
import os
from typing import Any, Dict
import pandas as pd
from llm_studio.src.datasets.conversation_chain_handler import get_conversation_chains
from llm_studio.src.datasets.text_utils import get_tokenizer
from llm_studio.src.utils.data_utils import read_dataframe_drop_missing_labels
from llm_studio.src.utils.plot_utils import (
PlotData,
format_for_markdown_visualization,
list_to_markdown_representation,
)
class Plots:
@classmethod
def plot_batch(cls, batch, cfg) -> PlotData:
tokenizer = get_tokenizer(cfg)
df = create_batch_prediction_df(batch, tokenizer)
path = os.path.join(cfg.output_directory, "batch_viz.parquet")
df.to_parquet(path)
return PlotData(path, encoding="df")
@classmethod
def plot_data(cls, cfg) -> PlotData:
"""
Plots the data in a scrollable table.
We limit the number of rows to max 600 to avoid rendering issues in Wave.
As the data visualization is instantiated on every page load, we cache the
data visualization in a parquet file.
"""
config_id = (
str(cfg.dataset.train_dataframe)
+ str(cfg.dataset.system_column)
+ str(cfg.dataset.prompt_column)
+ str(cfg.dataset.answer_column)
+ str(cfg.dataset.parent_id_column)
)
config_hash = hashlib.md5(config_id.encode()).hexdigest()
path = os.path.join(
os.path.dirname(cfg.dataset.train_dataframe),
f"__meta_info__{config_hash}_data_viz.parquet",
)
if os.path.exists(path):
return PlotData(path, encoding="df")
df = read_dataframe_drop_missing_labels(cfg.dataset.train_dataframe, cfg)
conversations = get_conversation_chains(df, cfg, limit_chained_samples=True)
# Limit to max 15 prompt-conversation-answer rounds
# This yields to max 5 * sum_{i=1}^{15} i = 600 rows in the DataFrame
max_conversation_length = min(
max([len(conversation["prompts"]) for conversation in conversations]), 15
)
conversations_to_display = []
for conversation_length in range(1, max_conversation_length + 1):
conversations_to_display += [
conversation
for conversation in conversations
if len(conversation["prompts"]) == conversation_length
][:5]
# Convert into a scrollable table by transposing the dataframe
df_transposed = pd.DataFrame(columns=["Sample Number", "Field", "Content"])
i = 0
for sample_number, conversation in enumerate(conversations_to_display):
if conversation["systems"][0] != "":
df_transposed.loc[i] = [
sample_number,
"System",
conversation["systems"][0],
]
i += 1
for prompt, answer in zip(conversation["prompts"], conversation["answers"]):
df_transposed.loc[i] = [
sample_number,
"Prompt",
prompt,
]
i += 1
df_transposed.loc[i] = [
sample_number,
"Answer",
answer,
]
i += 1
df_transposed["Content"] = df_transposed["Content"].apply(
format_for_markdown_visualization
)
df_transposed.to_parquet(path)
return PlotData(path, encoding="df")
@classmethod
def plot_validation_predictions(
cls, val_outputs: Dict, cfg: Any, val_df: pd.DataFrame, mode: str
) -> PlotData:
return plot_validation_predictions(val_outputs, cfg, val_df, mode)
def plot_validation_predictions(
val_outputs: Dict, cfg: Any, val_df: pd.DataFrame, mode: str
) -> PlotData:
conversations = get_conversation_chains(
val_df, cfg, limit_chained_samples=cfg.dataset.limit_chained_samples
)
prompt_column_name = (
cfg.dataset.prompt_column
if len(cfg.dataset.prompt_column) > 1
else cfg.dataset.prompt_column[0]
)
target_texts = [conversation["answers"][-1] for conversation in conversations]
input_texts = []
for conversation in conversations:
input_text = conversation["systems"][0]
prompts = conversation["prompts"]
answers = conversation["answers"]
# exclude last answer
answers[-1] = ""
for prompt, answer in zip(prompts, answers):
input_text += (
f" **{prompt_column_name}:** "
f"{prompt}\n\n"
f"**{cfg.dataset.answer_column}:** "
f"{answer}\n\n"
)
input_texts += [input_text]
if "predicted_text" in val_outputs.keys():
predicted_texts = val_outputs["predicted_text"]
else:
predicted_texts = [
"No predictions are generated for the selected metric"
] * len(target_texts)
input_text_column_name = (
"Input Text (tokenization max length setting "
"may truncate the input text during training/inference)"
)
df = pd.DataFrame(
{
input_text_column_name: input_texts,
"Target Text": target_texts,
"Predicted Text": predicted_texts,
}
)
df[input_text_column_name] = df[input_text_column_name].apply(
format_for_markdown_visualization
)
df["Target Text"] = df["Target Text"].apply(format_for_markdown_visualization)
df["Predicted Text"] = df["Predicted Text"].apply(format_for_markdown_visualization)
if val_outputs.get("metrics") is not None:
metric_column_name = f"Metric ({cfg.prediction.metric})"
df[metric_column_name] = val_outputs["metrics"]
df[metric_column_name] = df[metric_column_name].round(decimals=3)
if len(df) > 900:
df.sort_values(by=metric_column_name, inplace=True)
df = pd.concat(
[
df.iloc[:300],
df.iloc[300:-300].sample(n=300, random_state=42),
df.iloc[-300:],
]
).reset_index(drop=True)
elif len(df) > 900:
df = df.sample(n=900, random_state=42).reset_index(drop=True)
if val_outputs.get("explanations") is not None:
df["Explanation"] = val_outputs["explanations"]
path = os.path.join(cfg.output_directory, f"{mode}_viz.parquet")
df.to_parquet(path)
return PlotData(data=path, encoding="df")
def create_batch_prediction_df(
batch, tokenizer, ids_for_tokenized_text="input_ids", labels_column="labels"
):
df = pd.DataFrame(
{
"Prompt Text": [
tokenizer.decode(input_ids, skip_special_tokens=True)
for input_ids in batch["prompt_input_ids"].detach().cpu().numpy()
]
}
)
df["Prompt Text"] = df["Prompt Text"].apply(format_for_markdown_visualization)
if labels_column in batch.keys():
df["Answer Text"] = [
tokenizer.decode(
[label for label in labels if label != -100],
skip_special_tokens=True,
)
for labels in batch.get(labels_column, batch[ids_for_tokenized_text])
.detach()
.cpu()
.numpy()
]
tokens_list = [
tokenizer.convert_ids_to_tokens(input_ids)
for input_ids in batch[ids_for_tokenized_text].detach().cpu().numpy()
]
masks_list = [
[label != -100 for label in labels]
for labels in batch.get(labels_column, batch[ids_for_tokenized_text])
.detach()
.cpu()
.numpy()
]
df["Tokenized Text"] = [
list_to_markdown_representation(
tokens, masks, pad_token=tokenizer.pad_token, num_chars=100
)
for tokens, masks in zip(tokens_list, masks_list)
]
# limit to 2000 rows, still renders fast in wave
df = df.iloc[:2000]
# Convert into a scrollable table by transposing the dataframe
df_transposed = pd.DataFrame(columns=["Sample Number", "Field", "Content"])
has_answer = "Answer Text" in df.columns
for i, row in df.iterrows():
offset = 2 + int(has_answer)
df_transposed.loc[i * offset] = [
i,
"Prompt Text",
row["Prompt Text"],
]
if has_answer:
df_transposed.loc[i * offset + 1] = [
i,
"Answer Text",
row["Answer Text"],
]
df_transposed.loc[i * offset + 1 + int(has_answer)] = [
i,
"Tokenized Text",
row["Tokenized Text"],
]
return df_transposed
|