Spaces:
Running
Running
zamalali
commited on
Commit
Β·
9a94c10
0
Parent(s):
Fresh start without binary files
Browse files- .gitattributes +2 -0
- app.py +381 -0
- finetune_augmentor/__init__.py +1 -0
- finetune_augmentor/augmentor.py +583 -0
- requirements.txt +6 -0
.gitattributes
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
2 |
+
*.pdf filter=lfs diff=lfs merge=lfs -text
|
app.py
ADDED
@@ -0,0 +1,381 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
from finetune_augmentor import AugmentationExample, AugmentationConfig, FinetuningDataAugmentor
|
3 |
+
import json
|
4 |
+
import streamlit.components.v1 as components
|
5 |
+
from streamlit_ace import st_ace # Editable code block
|
6 |
+
|
7 |
+
# -------------------------------
|
8 |
+
# Page Configuration and CSS
|
9 |
+
# -------------------------------
|
10 |
+
|
11 |
+
st.set_page_config(
|
12 |
+
page_title="Finetuning Data Augmentation Generator",
|
13 |
+
layout="wide",
|
14 |
+
initial_sidebar_state="expanded",
|
15 |
+
)
|
16 |
+
|
17 |
+
|
18 |
+
components.html(
|
19 |
+
"""
|
20 |
+
|
21 |
+
<div style="position: fixed; top: 10px; right: 10px; z-index: 100;">
|
22 |
+
<a href="https://github.com/zamalali/ftboost" target="_blank">
|
23 |
+
<img src="https://github.githubassets.com/images/modules/logos_page/GitHub-Mark.png" alt="GitHub" style="height: 30px; margin-right: 10px;">
|
24 |
+
</a>
|
25 |
+
<a href="https://huggingface.co/zamal" target="_blank">
|
26 |
+
<img src="https://huggingface.co/front/assets/huggingface_logo.svg" alt="Hugging Face" style="height: 30px;">
|
27 |
+
</a>
|
28 |
+
</div>
|
29 |
+
""",
|
30 |
+
height=40
|
31 |
+
)
|
32 |
+
|
33 |
+
|
34 |
+
st.markdown(
|
35 |
+
"""
|
36 |
+
<style>
|
37 |
+
/* Main content area */
|
38 |
+
.block-container {
|
39 |
+
background-color: #121212;
|
40 |
+
color: #ffffff;
|
41 |
+
}
|
42 |
+
/* Sidebar styling */
|
43 |
+
[data-testid="stSidebar"] {
|
44 |
+
background-color: #121212;
|
45 |
+
color: #ffffff;
|
46 |
+
}
|
47 |
+
[data-testid="stSidebar"] * {
|
48 |
+
color: #ffffff !important;
|
49 |
+
}
|
50 |
+
/* Button styling */
|
51 |
+
.stButton>button, .stDownloadButton>button {
|
52 |
+
background-color: #808080 !important;
|
53 |
+
color: #ffffff !important;
|
54 |
+
font-size: 16px;
|
55 |
+
border: none;
|
56 |
+
border-radius: 5px;
|
57 |
+
padding: 0.5rem 1.5rem;
|
58 |
+
margin-top: 1rem;
|
59 |
+
}
|
60 |
+
/* Text inputs */
|
61 |
+
.stTextInput>div>input, .stNumberInput>div>input {
|
62 |
+
border-radius: 5px;
|
63 |
+
border: 1px solid #ffffff;
|
64 |
+
padding: 0.5rem;
|
65 |
+
background-color: #1a1a1a;
|
66 |
+
color: #ffffff;
|
67 |
+
}
|
68 |
+
.stTextArea>textarea {
|
69 |
+
background-color: #1a1a1a;
|
70 |
+
color: #ffffff;
|
71 |
+
font-family: "Courier New", monospace;
|
72 |
+
border: 1px solid #ffffff;
|
73 |
+
border-radius: 5px;
|
74 |
+
padding: 1rem;
|
75 |
+
}
|
76 |
+
/* Header colors */
|
77 |
+
h1 { color: #00FF00; }
|
78 |
+
h2, h3, h4 { color: #FFFF00; }
|
79 |
+
/* Field labels */
|
80 |
+
label { color: #ffffff !important; }
|
81 |
+
/* Remove extra margin in code blocks */
|
82 |
+
pre { margin: 0; }
|
83 |
+
/* Ace editor style overrides */
|
84 |
+
.ace_editor {
|
85 |
+
border: none !important;
|
86 |
+
box-shadow: none !important;
|
87 |
+
background-color: #121212 !important;
|
88 |
+
}
|
89 |
+
/* Override alert (error/success) text colors */
|
90 |
+
[data-testid="stAlert"] { color: #ffffff !important; }
|
91 |
+
/* Add white border to expander header */
|
92 |
+
[data-testid="stExpander"] > div:first-child {
|
93 |
+
border: 1px solid #ffffff !important;
|
94 |
+
}
|
95 |
+
</style>
|
96 |
+
""",
|
97 |
+
unsafe_allow_html=True,
|
98 |
+
)
|
99 |
+
|
100 |
+
# Inject JavaScript to scroll to top on load
|
101 |
+
components.html(
|
102 |
+
"""
|
103 |
+
<script>
|
104 |
+
document.addEventListener("DOMContentLoaded", function() {
|
105 |
+
setTimeout(function() { window.scrollTo(0, 0); }, 100);
|
106 |
+
});
|
107 |
+
</script>
|
108 |
+
""",
|
109 |
+
height=0,
|
110 |
+
)
|
111 |
+
|
112 |
+
# -------------------------------
|
113 |
+
# App Title and Description
|
114 |
+
# -------------------------------
|
115 |
+
st.title("ftBoost π")
|
116 |
+
st.markdown(
|
117 |
+
"""
|
118 |
+
**ftBoost Hero** is a powerful tool designed to help you generate high-quality fine-tuning data for AI models.
|
119 |
+
Whether you're working with OpenAI, Gemini, Mistral, or LLaMA models, this app allows you to create structured
|
120 |
+
input-output pairs and apply augmentation techniques to enhance dataset quality. With advanced tuning parameters,
|
121 |
+
semantic similarity controls, and fluency optimization, **ftBoost Hero** ensures that your fine-tuning data is diverse,
|
122 |
+
well-structured, and ready for training. π
|
123 |
+
""",
|
124 |
+
unsafe_allow_html=True,
|
125 |
+
)
|
126 |
+
|
127 |
+
# -------------------------------
|
128 |
+
# Step A: File Upload & Auto-Detection
|
129 |
+
# -------------------------------
|
130 |
+
st.markdown("##### Step 1: Upload Your Finetuning data JSONL File if you have one already (Optional)")
|
131 |
+
uploaded_file = st.file_uploader("Upload your train.jsonl file", type=["jsonl", "txt"])
|
132 |
+
uploaded_examples = []
|
133 |
+
detected_model = None
|
134 |
+
|
135 |
+
if uploaded_file is not None:
|
136 |
+
try:
|
137 |
+
file_content = uploaded_file.getvalue().decode("utf-8")
|
138 |
+
# Auto-detect model type from the first valid snippet
|
139 |
+
for line in file_content.splitlines():
|
140 |
+
if line.strip():
|
141 |
+
record = json.loads(line)
|
142 |
+
if "messages" in record:
|
143 |
+
msgs = record["messages"]
|
144 |
+
if len(msgs) >= 3 and msgs[0].get("role") == "system":
|
145 |
+
detected_model = "OpenAI Models"
|
146 |
+
elif len(msgs) == 2:
|
147 |
+
detected_model = "Mistral Models"
|
148 |
+
elif "contents" in record:
|
149 |
+
detected_model = "Gemini Models"
|
150 |
+
break
|
151 |
+
|
152 |
+
# Display an info message based on detection result
|
153 |
+
if detected_model is not None:
|
154 |
+
st.info(f"This JSONL file format supports the **{detected_model}**.")
|
155 |
+
else:
|
156 |
+
st.info("The uploaded JSONL file format is not recognized. Please manually select the appropriate model.")
|
157 |
+
|
158 |
+
# Process the entire file for valid examples
|
159 |
+
for line in file_content.splitlines():
|
160 |
+
if not line.strip():
|
161 |
+
continue
|
162 |
+
record = json.loads(line)
|
163 |
+
input_text, output_text = "", ""
|
164 |
+
if "messages" in record:
|
165 |
+
msgs = record["messages"]
|
166 |
+
if len(msgs) >= 3:
|
167 |
+
input_text = msgs[1].get("content", "").strip()
|
168 |
+
output_text = msgs[2].get("content", "").strip()
|
169 |
+
elif len(msgs) == 2:
|
170 |
+
input_text = msgs[0].get("content", "").strip()
|
171 |
+
output_text = msgs[1].get("content", "").strip()
|
172 |
+
elif "contents" in record:
|
173 |
+
contents = record["contents"]
|
174 |
+
if len(contents) >= 2 and "parts" in contents[0] and "parts" in contents[1]:
|
175 |
+
input_text = contents[0]["parts"][0].get("text", "").strip()
|
176 |
+
output_text = contents[1]["parts"][0].get("text", "").strip()
|
177 |
+
if input_text and output_text:
|
178 |
+
uploaded_examples.append(AugmentationExample(input_text=input_text, output_text=output_text))
|
179 |
+
if len(uploaded_examples) < 3:
|
180 |
+
st.error("Uploaded file does not contain at least 3 valid input/output pairs.")
|
181 |
+
else:
|
182 |
+
st.success(f"Uploaded file processed: {len(uploaded_examples)} valid input/output pairs loaded.")
|
183 |
+
except Exception as e:
|
184 |
+
st.error(f"Error processing uploaded file: {e}")
|
185 |
+
|
186 |
+
# -------------------------------
|
187 |
+
# Step B: Model Selection
|
188 |
+
# -------------------------------
|
189 |
+
default_model = detected_model if detected_model is not None else "OpenAI Models"
|
190 |
+
model_options = ["OpenAI Models", "Gemini Models", "Mistral Models", "Llama Models"]
|
191 |
+
default_index = model_options.index(default_model) if default_model in model_options else 0
|
192 |
+
model_type = st.selectbox(
|
193 |
+
"Select the output format for finetuning",
|
194 |
+
model_options,
|
195 |
+
index=default_index
|
196 |
+
)
|
197 |
+
|
198 |
+
# -------------------------------
|
199 |
+
# Step C: System Message & API Key
|
200 |
+
# -------------------------------
|
201 |
+
system_message = st.text_input("System Message (optional) only for OpenAI models", value="Marv is a factual chatbot that is also sarcastic.")
|
202 |
+
# groq_api_key = st.text_input("LangChain Groq API Key", type="password", help="Enter your LangChain Groq API Key for data augmentation")
|
203 |
+
|
204 |
+
|
205 |
+
|
206 |
+
groq_api_key = st.text_input(
|
207 |
+
"LangChain Groq API Key (if you don't have one, get it from [here](https://console.groq.com/keys))",
|
208 |
+
type="password",
|
209 |
+
help="Enter your LangChain Groq API Key for data augmentation"
|
210 |
+
)
|
211 |
+
# -------------------------------
|
212 |
+
# Step D: Input Schema Template Display
|
213 |
+
# -------------------------------
|
214 |
+
st.markdown("#### Input Schema Template")
|
215 |
+
if model_type == "OpenAI Models":
|
216 |
+
st.code(
|
217 |
+
'''{
|
218 |
+
"messages": [
|
219 |
+
{"role": "system", "content": "Marv is a factual chatbot that is also sarcastic."},
|
220 |
+
{"role": "user", "content": "What's the capital of France?"},
|
221 |
+
{"role": "assistant", "content": "Paris, as if everyone doesn't know that already."}
|
222 |
+
]
|
223 |
+
}''', language="json")
|
224 |
+
elif model_type == "Gemini Models":
|
225 |
+
st.code(
|
226 |
+
'''{
|
227 |
+
"contents": [
|
228 |
+
{"role": "user", "parts": [{"text": "What's the capital of France?"}]},
|
229 |
+
{"role": "model", "parts": [{"text": "Paris, as if everyone doesn't know that already."}]}
|
230 |
+
]
|
231 |
+
}''', language="json")
|
232 |
+
else:
|
233 |
+
st.code(
|
234 |
+
'''{
|
235 |
+
"messages": [
|
236 |
+
{"role": "user", "content": "What's the capital of France?"},
|
237 |
+
{"role": "assistant", "content": "Paris, as if everyone doesn't know that already."}
|
238 |
+
]
|
239 |
+
}''', language="json")
|
240 |
+
|
241 |
+
# -------------------------------
|
242 |
+
# Step E: Manual Input of Pairs (if no file uploaded)
|
243 |
+
# -------------------------------
|
244 |
+
if uploaded_file is None:
|
245 |
+
st.markdown("##### Enter at least 3 input/output pairs manually:")
|
246 |
+
num_pairs = st.number_input("Number of Pairs", min_value=3, value=3, step=1)
|
247 |
+
pair_templates = []
|
248 |
+
for i in range(num_pairs):
|
249 |
+
st.markdown(f"##### Pair {i+1}")
|
250 |
+
if model_type == "OpenAI Models":
|
251 |
+
init_template = ('''{
|
252 |
+
"messages": [
|
253 |
+
{"role": "system", "content": "''' + system_message + '''"},
|
254 |
+
{"role": "user", "content": "Enter your input text here"},
|
255 |
+
{"role": "assistant", "content": "Enter your output text here"}
|
256 |
+
]
|
257 |
+
}''').strip()
|
258 |
+
ace_key = f"pair_{i}_{model_type}_{system_message}"
|
259 |
+
elif model_type == "Gemini Models":
|
260 |
+
init_template = ('''{
|
261 |
+
"contents": [
|
262 |
+
{"role": "user", "parts": [{"text": "Enter your input text here"}]},
|
263 |
+
{"role": "model", "parts": [{"text": "Enter your output text here"}]}
|
264 |
+
]
|
265 |
+
}''').strip()
|
266 |
+
ace_key = f"pair_{i}_{model_type}"
|
267 |
+
else:
|
268 |
+
init_template = ('''{
|
269 |
+
"messages": [
|
270 |
+
{"role": "user", "content": "Enter your input text here"},
|
271 |
+
{"role": "assistant", "content": "Enter your output text here"}
|
272 |
+
]
|
273 |
+
}''').strip()
|
274 |
+
ace_key = f"pair_{i}_{model_type}"
|
275 |
+
|
276 |
+
pair = st_ace(
|
277 |
+
placeholder="Edit your code here...",
|
278 |
+
value=init_template,
|
279 |
+
language="json",
|
280 |
+
theme="monokai",
|
281 |
+
key=ace_key,
|
282 |
+
height=150
|
283 |
+
)
|
284 |
+
pair_templates.append(pair)
|
285 |
+
|
286 |
+
# -------------------------------
|
287 |
+
# Step F: Augmentation Settings
|
288 |
+
# -------------------------------
|
289 |
+
target_augmented = st.number_input("Number of Augmented Pairs to Generate", min_value=5, value=5, step=1)
|
290 |
+
finetuning_goal = "Improve conversational clarity and capture subtle nuances"
|
291 |
+
st.markdown(f"**Finetuning Goal:** {finetuning_goal}")
|
292 |
+
|
293 |
+
with st.expander("Show/Hide Advanced Tuning Parameters"):
|
294 |
+
min_semantic = st.slider("Minimum Semantic Similarity", 0.0, 1.0, 0.80, 0.01)
|
295 |
+
max_semantic = st.slider("Maximum Semantic Similarity", 0.0, 1.0, 0.95, 0.01)
|
296 |
+
min_diversity = st.slider("Minimum Diversity Score", 0.0, 1.0, 0.70, 0.01)
|
297 |
+
min_fluency = st.slider("Minimum Fluency Score", 0.0, 1.0, 0.80, 0.01)
|
298 |
+
|
299 |
+
# -------------------------------
|
300 |
+
# Step G: Generate Data Button and Pipeline Execution
|
301 |
+
# -------------------------------
|
302 |
+
if st.button("Generate Data"):
|
303 |
+
if not groq_api_key.strip():
|
304 |
+
st.error("Please enter your LangChain Groq API Key to proceed.")
|
305 |
+
st.stop()
|
306 |
+
|
307 |
+
# Choose examples: from uploaded file if available; otherwise from manual input.
|
308 |
+
if uploaded_file is not None and len(uploaded_examples) >= 3:
|
309 |
+
examples = uploaded_examples
|
310 |
+
else:
|
311 |
+
examples = []
|
312 |
+
errors = []
|
313 |
+
for idx, pair in enumerate(pair_templates):
|
314 |
+
try:
|
315 |
+
record = json.loads(pair)
|
316 |
+
if model_type == "OpenAI Models":
|
317 |
+
msgs = record.get("messages", [])
|
318 |
+
if len(msgs) != 3:
|
319 |
+
raise ValueError("Expected 3 messages")
|
320 |
+
input_text = msgs[1].get("content", "").strip()
|
321 |
+
output_text = msgs[2].get("content", "").strip()
|
322 |
+
elif model_type == "Gemini Models":
|
323 |
+
contents = record.get("contents", [])
|
324 |
+
if len(contents) < 2:
|
325 |
+
raise ValueError("Expected at least 2 contents")
|
326 |
+
input_text = contents[0]["parts"][0].get("text", "").strip()
|
327 |
+
output_text = contents[1]["parts"][0].get("text", "").strip()
|
328 |
+
else:
|
329 |
+
msgs = record.get("messages", [])
|
330 |
+
if len(msgs) != 2:
|
331 |
+
raise ValueError("Expected 2 messages for this format")
|
332 |
+
input_text = msgs[0].get("content", "").strip()
|
333 |
+
output_text = msgs[1].get("content", "").strip()
|
334 |
+
if not input_text or not output_text:
|
335 |
+
raise ValueError("Input or output text is empty")
|
336 |
+
examples.append(AugmentationExample(input_text=input_text, output_text=output_text))
|
337 |
+
except Exception as e:
|
338 |
+
errors.append(f"Error in pair {idx+1}: {e}")
|
339 |
+
if errors:
|
340 |
+
st.error("There were errors in your input pairs:\n" + "\n".join(errors))
|
341 |
+
elif len(examples) < 3:
|
342 |
+
st.error("Please provide at least 3 valid pairs.")
|
343 |
+
|
344 |
+
if len(examples) >= 3:
|
345 |
+
target_model = "mixtral-8x7b-32768"
|
346 |
+
try:
|
347 |
+
config = AugmentationConfig(
|
348 |
+
target_model=target_model,
|
349 |
+
examples=examples,
|
350 |
+
finetuning_goal=finetuning_goal,
|
351 |
+
groq_api_key=groq_api_key,
|
352 |
+
system_message=system_message,
|
353 |
+
min_semantic_similarity=min_semantic,
|
354 |
+
max_semantic_similarity=max_semantic,
|
355 |
+
min_diversity_score=min_diversity,
|
356 |
+
min_fluency_score=min_fluency
|
357 |
+
)
|
358 |
+
except Exception as e:
|
359 |
+
st.error(f"Configuration error: {e}")
|
360 |
+
st.stop()
|
361 |
+
|
362 |
+
st.markdown('<p style="color: white;">Running augmentation pipeline... Please wait.</p>', unsafe_allow_html=True)
|
363 |
+
|
364 |
+
augmentor = FinetuningDataAugmentor(config)
|
365 |
+
augmentor.run_augmentation(target_count=target_augmented)
|
366 |
+
|
367 |
+
fmt = model_type.lower()
|
368 |
+
if fmt == "openai models":
|
369 |
+
output_data = augmentor.get_formatted_output(format_type="openai")
|
370 |
+
elif fmt == "gemini models":
|
371 |
+
output_data = augmentor.get_formatted_output(format_type="gemini")
|
372 |
+
elif fmt == "mistral models":
|
373 |
+
output_data = augmentor.get_formatted_output(format_type="mistral")
|
374 |
+
elif fmt == "llama models":
|
375 |
+
output_data = augmentor.get_formatted_output(format_type="llama")
|
376 |
+
else:
|
377 |
+
output_data = augmentor.get_formatted_output(format_type="openai")
|
378 |
+
|
379 |
+
st.markdown("### Augmented Data")
|
380 |
+
st.code(output_data, language="json")
|
381 |
+
st.download_button("Download train.jsonl", output_data, file_name="train.jsonl")
|
finetune_augmentor/__init__.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
from .augmentor import AugmentationExample, AugmentationConfig, FinetuningDataAugmentor, load_examples_from_file
|
finetune_augmentor/augmentor.py
ADDED
@@ -0,0 +1,583 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
augmentor.py
|
3 |
+
|
4 |
+
This module implements a robust and scalable pipeline for finetuning data augmentation.
|
5 |
+
It supports generating augmented data in either OpenAI, Gemini, Mistral, or LLama fineβtuning JSONL format.
|
6 |
+
Users may optionally override metric thresholds and load existing examples from a JSONL file.
|
7 |
+
The LangChain Groq API key is now provided via the configuration rather than the .env file.
|
8 |
+
"""
|
9 |
+
|
10 |
+
import os
|
11 |
+
import json
|
12 |
+
import uuid
|
13 |
+
import logging
|
14 |
+
import re
|
15 |
+
import random
|
16 |
+
import ast
|
17 |
+
from typing import List, Dict, Any, Optional
|
18 |
+
|
19 |
+
# Removed dotenv load for GROQ_API_KEY since it is now provided in config
|
20 |
+
# Configure logging
|
21 |
+
logging.basicConfig(level=logging.INFO)
|
22 |
+
logger = logging.getLogger("FinetuningAugmentor")
|
23 |
+
|
24 |
+
# Environment tokens (kept for HF_TOKEN if needed)
|
25 |
+
HF_TOKEN = os.getenv("HF_TOKEN")
|
26 |
+
# GROQ_API_KEY will now be provided in the configuration
|
27 |
+
|
28 |
+
# -----------------------------
|
29 |
+
# Data Models and Preprocessing
|
30 |
+
# -----------------------------
|
31 |
+
from pydantic import BaseModel, field_validator, ValidationError
|
32 |
+
|
33 |
+
class AugmentationExample(BaseModel):
|
34 |
+
"""
|
35 |
+
An input/output example for augmentation.
|
36 |
+
"""
|
37 |
+
input_text: str
|
38 |
+
output_text: str
|
39 |
+
|
40 |
+
@field_validator('input_text', 'output_text')
|
41 |
+
def non_empty(cls, v: str) -> str:
|
42 |
+
if not v.strip():
|
43 |
+
raise ValueError("Text fields must be non-empty")
|
44 |
+
return v.strip()
|
45 |
+
|
46 |
+
class AugmentationConfig(BaseModel):
|
47 |
+
"""
|
48 |
+
Configuration for the augmentation process.
|
49 |
+
"""
|
50 |
+
target_model: str # e.g., "mixtral-8x7b-32768" or any Groq-supported model name
|
51 |
+
examples: List[AugmentationExample]
|
52 |
+
finetuning_goal: str
|
53 |
+
groq_api_key: str
|
54 |
+
system_message: Optional[str] = "Marv is a factual chatbot that is also sarcastic."
|
55 |
+
# Optional metric thresholds (if not provided, defaults are used)
|
56 |
+
min_semantic_similarity: Optional[float] = 0.80
|
57 |
+
max_semantic_similarity: Optional[float] = 0.95
|
58 |
+
min_diversity_score: Optional[float] = 0.70
|
59 |
+
min_fluency_score: Optional[float] = 0.80
|
60 |
+
|
61 |
+
@field_validator('examples')
|
62 |
+
def check_examples_length(cls, v: List[AugmentationExample]) -> List[AugmentationExample]:
|
63 |
+
if len(v) < 3:
|
64 |
+
raise ValueError("Provide at least 3 examples")
|
65 |
+
return v
|
66 |
+
|
67 |
+
class StandardExample(BaseModel):
|
68 |
+
"""
|
69 |
+
Standardized format for input examples.
|
70 |
+
"""
|
71 |
+
id: str
|
72 |
+
input_text: str
|
73 |
+
output_text: str
|
74 |
+
metadata: Dict[str, Any] = {}
|
75 |
+
|
76 |
+
def normalize_examples(examples: List[AugmentationExample]) -> List[StandardExample]:
|
77 |
+
"""
|
78 |
+
Normalize and standardize input examples.
|
79 |
+
"""
|
80 |
+
normalized = []
|
81 |
+
for ex in examples:
|
82 |
+
norm_ex = StandardExample(
|
83 |
+
id=str(uuid.uuid4()),
|
84 |
+
input_text=ex.input_text.lower(),
|
85 |
+
output_text=ex.output_text.lower(),
|
86 |
+
metadata={"original_word_count": len(ex.input_text.split())}
|
87 |
+
)
|
88 |
+
normalized.append(norm_ex)
|
89 |
+
logger.info(f"Normalized {len(normalized)} examples.")
|
90 |
+
return normalized
|
91 |
+
|
92 |
+
# -----------------------------
|
93 |
+
# Dynamic Strategy Selection
|
94 |
+
# -----------------------------
|
95 |
+
def determine_augmentation_strategy(config: AugmentationConfig) -> Dict[str, Any]:
|
96 |
+
"""
|
97 |
+
Determine the augmentation strategy based on the finetuning goal.
|
98 |
+
"""
|
99 |
+
goal = config.finetuning_goal.lower()
|
100 |
+
strategy = {}
|
101 |
+
if any(word in goal for word in ["dialogue", "q&a", "conversation", "chat"]):
|
102 |
+
strategy["methods"] = ["llm_paraphrasing", "back_translation"]
|
103 |
+
else:
|
104 |
+
strategy["methods"] = ["eda_synonym_replacement", "llm_paraphrasing", "synthetic_noise"]
|
105 |
+
strategy["diversity_threshold"] = 0.7
|
106 |
+
logger.info(f"Determined augmentation strategy: {strategy}")
|
107 |
+
return strategy
|
108 |
+
|
109 |
+
# -----------------------------
|
110 |
+
# Helper Functions
|
111 |
+
# -----------------------------
|
112 |
+
def extract_json(text: str) -> dict:
|
113 |
+
"""
|
114 |
+
Extract the first valid JSON object from a given text.
|
115 |
+
"""
|
116 |
+
json_pattern = re.compile(r'\{.*\}', re.DOTALL)
|
117 |
+
match = json_pattern.search(text)
|
118 |
+
if match:
|
119 |
+
json_str = match.group()
|
120 |
+
try:
|
121 |
+
return json.loads(json_str)
|
122 |
+
except json.JSONDecodeError as e:
|
123 |
+
raise ValueError(f"JSON decoding error: {e}")
|
124 |
+
else:
|
125 |
+
raise ValueError("No valid JSON found in text.")
|
126 |
+
|
127 |
+
def make_hashable(item: Any) -> Any:
|
128 |
+
"""
|
129 |
+
Recursively convert unhashable types (lists/dicts) into hashable tuples.
|
130 |
+
"""
|
131 |
+
if isinstance(item, (list, tuple)):
|
132 |
+
return tuple(make_hashable(i) for i in item)
|
133 |
+
elif isinstance(item, dict):
|
134 |
+
return tuple(sorted((k, make_hashable(v)) for k, v in item.items()))
|
135 |
+
else:
|
136 |
+
return item
|
137 |
+
|
138 |
+
def validate_jsonl_record(record: dict) -> bool:
|
139 |
+
"""
|
140 |
+
Validates that the record follows the required OpenAI format:
|
141 |
+
{"messages": [{"role": "system", "content": <str>},
|
142 |
+
{"role": "user", "content": <non-empty str>},
|
143 |
+
{"role": "assistant", "content": <non-empty str>}]}
|
144 |
+
"""
|
145 |
+
if "messages" not in record:
|
146 |
+
logger.error("Record missing 'messages' key.")
|
147 |
+
return False
|
148 |
+
messages = record["messages"]
|
149 |
+
if not isinstance(messages, list) or len(messages) != 3:
|
150 |
+
logger.error("Record 'messages' must be a list of 3 items.")
|
151 |
+
return False
|
152 |
+
expected_roles = ["system", "user", "assistant"]
|
153 |
+
for msg, role in zip(messages, expected_roles):
|
154 |
+
if not isinstance(msg, dict):
|
155 |
+
logger.error("Each message must be a dictionary.")
|
156 |
+
return False
|
157 |
+
if msg.get("role") != role:
|
158 |
+
logger.error(f"Expected role '{role}', but got '{msg.get('role')}'.")
|
159 |
+
return False
|
160 |
+
if "content" not in msg or not isinstance(msg["content"], str):
|
161 |
+
logger.error("Each message must have a string 'content' field.")
|
162 |
+
return False
|
163 |
+
if role in ["user", "assistant"] and not msg["content"].strip():
|
164 |
+
logger.error(f"Message for role '{role}' has empty content.")
|
165 |
+
return False
|
166 |
+
return True
|
167 |
+
|
168 |
+
def get_text(value: Any) -> str:
|
169 |
+
"""
|
170 |
+
Ensure the value is returned as a string.
|
171 |
+
If it is a list, recursively return the first element.
|
172 |
+
If it is a dict and contains a "text" key, return that.
|
173 |
+
If it is a string that resembles a dict, try to parse it.
|
174 |
+
"""
|
175 |
+
if isinstance(value, list):
|
176 |
+
if value:
|
177 |
+
return get_text(value[0])
|
178 |
+
return ""
|
179 |
+
elif isinstance(value, dict):
|
180 |
+
if "text" in value:
|
181 |
+
return str(value["text"])
|
182 |
+
return str(value)
|
183 |
+
elif isinstance(value, str):
|
184 |
+
val = value.strip()
|
185 |
+
if val.startswith("{") and val.endswith("}"):
|
186 |
+
try:
|
187 |
+
parsed = ast.literal_eval(val)
|
188 |
+
if isinstance(parsed, dict) and "text" in parsed:
|
189 |
+
return str(parsed["text"])
|
190 |
+
except Exception:
|
191 |
+
pass
|
192 |
+
return val
|
193 |
+
else:
|
194 |
+
return str(value)
|
195 |
+
|
196 |
+
# --- New helper: Fix content formatting ---
|
197 |
+
def fix_content(content: str) -> str:
|
198 |
+
"""
|
199 |
+
If the content appears to be a Python dict (using single quotes), try to
|
200 |
+
convert it to valid JSON (with double quotes). If parsing fails, return the original content.
|
201 |
+
"""
|
202 |
+
content = content.strip()
|
203 |
+
if content.startswith("{") and content.endswith("}") and "'" in content:
|
204 |
+
try:
|
205 |
+
parsed = ast.literal_eval(content)
|
206 |
+
return json.dumps(parsed)
|
207 |
+
except Exception as e:
|
208 |
+
logger.debug(f"Failed to fix content formatting: {e}")
|
209 |
+
return content
|
210 |
+
|
211 |
+
def flatten_content(content: str) -> str:
|
212 |
+
"""
|
213 |
+
If content (after fixing) is a JSON string representing a dictionary,
|
214 |
+
flatten it by joining its values into a single plain-text string.
|
215 |
+
"""
|
216 |
+
try:
|
217 |
+
parsed = json.loads(content)
|
218 |
+
if isinstance(parsed, dict):
|
219 |
+
# Join values in sorted order by key
|
220 |
+
values = [str(parsed[k]).strip() for k in sorted(parsed)]
|
221 |
+
return " ".join(values)
|
222 |
+
except Exception:
|
223 |
+
pass
|
224 |
+
return content
|
225 |
+
|
226 |
+
# -----------------------------
|
227 |
+
# Augmentation Generation via LangChain Groq
|
228 |
+
# -----------------------------
|
229 |
+
from langchain_groq import ChatGroq
|
230 |
+
from langchain_core.prompts import ChatPromptTemplate
|
231 |
+
|
232 |
+
def instantiate_groq_llm(model: str, groq_api_key: str) -> ChatGroq:
|
233 |
+
"""
|
234 |
+
Instantiate a ChatGroq LLM with the given model name and API key.
|
235 |
+
"""
|
236 |
+
return ChatGroq(
|
237 |
+
model=model,
|
238 |
+
temperature=0.7,
|
239 |
+
max_tokens=256,
|
240 |
+
timeout=30,
|
241 |
+
max_retries=2,
|
242 |
+
groq_api_key=groq_api_key
|
243 |
+
)
|
244 |
+
|
245 |
+
def generate_initial_augmentation(example: StandardExample,
|
246 |
+
config: AugmentationConfig,
|
247 |
+
strategy: Dict[str, Any]) -> dict:
|
248 |
+
"""
|
249 |
+
Generate an initial candidate augmentation using an LLM prompt chain.
|
250 |
+
"""
|
251 |
+
prompt_template = ChatPromptTemplate.from_messages([
|
252 |
+
(
|
253 |
+
"system",
|
254 |
+
("You are a creative augmentation assistant that produces diverse yet semantically consistent "
|
255 |
+
"input/output pairs for finetuning tasks.")
|
256 |
+
),
|
257 |
+
(
|
258 |
+
"human",
|
259 |
+
(
|
260 |
+
"Augment the following example using the methods: {methods}. The finetuning goal is: {finetuning_goal}.\n"
|
261 |
+
"Ensure your output is in valid JSON format with keys 'augmented_input' and 'augmented_output'.\n"
|
262 |
+
"Input: {input_text}\n"
|
263 |
+
"Output: {output_text}\n"
|
264 |
+
"Return only the JSON response."
|
265 |
+
)
|
266 |
+
)
|
267 |
+
])
|
268 |
+
prompt_vars = {
|
269 |
+
"methods": ", ".join(strategy["methods"]),
|
270 |
+
"finetuning_goal": config.finetuning_goal,
|
271 |
+
"input_text": example.input_text,
|
272 |
+
"output_text": example.output_text
|
273 |
+
}
|
274 |
+
chain = prompt_template | instantiate_groq_llm(config.target_model, config.groq_api_key)
|
275 |
+
ai_msg = chain.invoke(prompt_vars)
|
276 |
+
logger.info(f"Initial augmentation for {example.id}: {ai_msg.content.strip()}")
|
277 |
+
return extract_json(ai_msg.content.strip())
|
278 |
+
|
279 |
+
def refine_augmentation(candidate: dict,
|
280 |
+
example: StandardExample,
|
281 |
+
config: AugmentationConfig,
|
282 |
+
strategy: Dict[str, Any]) -> dict:
|
283 |
+
"""
|
284 |
+
Refine a candidate augmentation using a second LLM prompt chain.
|
285 |
+
"""
|
286 |
+
refinement_template = ChatPromptTemplate.from_messages([
|
287 |
+
(
|
288 |
+
"system",
|
289 |
+
"You are an expert data augmentation advisor who refines candidate augmentations to maximize semantic accuracy, diversity, and clarity."
|
290 |
+
),
|
291 |
+
(
|
292 |
+
"human",
|
293 |
+
(
|
294 |
+
"Review the candidate augmentation for the following input/output pair and refine it if needed.\n"
|
295 |
+
"Finetuning Goal: {finetuning_goal}\n"
|
296 |
+
"Original Input: {input_text}\n"
|
297 |
+
"Original Output: {output_text}\n"
|
298 |
+
"Candidate Augmentation: {candidate}\n"
|
299 |
+
"Return a refined augmentation in valid JSON format with keys 'augmented_input' and 'augmented_output' only."
|
300 |
+
)
|
301 |
+
)
|
302 |
+
])
|
303 |
+
refinement_vars = {
|
304 |
+
"finetuning_goal": config.finetuning_goal,
|
305 |
+
"input_text": example.input_text,
|
306 |
+
"output_text": example.output_text,
|
307 |
+
"candidate": json.dumps(candidate)
|
308 |
+
}
|
309 |
+
chain = refinement_template | instantiate_groq_llm(config.target_model, config.groq_api_key)
|
310 |
+
ai_msg = chain.invoke(refinement_vars)
|
311 |
+
try:
|
312 |
+
refined = extract_json(ai_msg.content.strip())
|
313 |
+
logger.info(f"Refined augmentation for {example.id}: {refined}")
|
314 |
+
return refined
|
315 |
+
except Exception as e:
|
316 |
+
logger.error(f"Refinement failed for {example.id}: {e}. Using original candidate.")
|
317 |
+
return candidate
|
318 |
+
|
319 |
+
def calculate_metrics(augmentation: dict, original: StandardExample) -> dict:
|
320 |
+
"""
|
321 |
+
Simulate metric calculations for the candidate augmentation.
|
322 |
+
"""
|
323 |
+
semantic_similarity = random.uniform(0.78, 0.97)
|
324 |
+
diversity_score = random.uniform(0.65, 0.9)
|
325 |
+
fluency_score = random.uniform(0.80, 0.95)
|
326 |
+
metrics = {
|
327 |
+
"semantic_similarity": semantic_similarity,
|
328 |
+
"diversity_score": diversity_score,
|
329 |
+
"fluency_score": fluency_score
|
330 |
+
}
|
331 |
+
logger.info(f"Metrics for candidate of {original.id}: {metrics}")
|
332 |
+
return metrics
|
333 |
+
|
334 |
+
def metrics_valid(metrics: dict, config: AugmentationConfig) -> bool:
|
335 |
+
"""
|
336 |
+
Validate metric thresholds using configuration values.
|
337 |
+
"""
|
338 |
+
if metrics["semantic_similarity"] < config.min_semantic_similarity or metrics["semantic_similarity"] > config.max_semantic_similarity:
|
339 |
+
return False
|
340 |
+
if metrics["diversity_score"] < config.min_diversity_score:
|
341 |
+
return False
|
342 |
+
if metrics["fluency_score"] < config.min_fluency_score:
|
343 |
+
return False
|
344 |
+
return True
|
345 |
+
|
346 |
+
def quality_check(augmentation: Dict[str, Any], config: AugmentationConfig) -> bool:
|
347 |
+
"""
|
348 |
+
Simulate an LLM-based QA check.
|
349 |
+
"""
|
350 |
+
qa_prompt = (
|
351 |
+
f"Verify that the following augmentation preserves the intended meaning and style for the input/output pair "
|
352 |
+
f"given the finetuning goal '{config.finetuning_goal}':\n"
|
353 |
+
f"{augmentation['augmentation']}\n"
|
354 |
+
"Answer 'yes' if valid, otherwise 'no'."
|
355 |
+
)
|
356 |
+
logger.debug(f"QA Prompt: {qa_prompt}")
|
357 |
+
return True # Simulation: always passes
|
358 |
+
|
359 |
+
def generate_augmentations(normalized_examples: List[StandardExample],
|
360 |
+
config: AugmentationConfig,
|
361 |
+
strategy: Dict[str, Any],
|
362 |
+
target_count: int = 50) -> List[Dict[str, Any]]:
|
363 |
+
"""
|
364 |
+
Repeatedly generate candidate augmentations until at least target_count valid candidates are collected.
|
365 |
+
"""
|
366 |
+
augmented_candidates = []
|
367 |
+
attempts = 0
|
368 |
+
max_attempts = 100 # Safety valve
|
369 |
+
while len(augmented_candidates) < target_count and attempts < max_attempts:
|
370 |
+
for ex in normalized_examples:
|
371 |
+
try:
|
372 |
+
candidate = generate_initial_augmentation(ex, config, strategy)
|
373 |
+
refined_candidate = refine_augmentation(candidate, ex, config, strategy)
|
374 |
+
metrics = calculate_metrics(refined_candidate, ex)
|
375 |
+
if not metrics_valid(metrics, config):
|
376 |
+
logger.info(f"Candidate for {ex.id} rejected by metrics: {metrics}")
|
377 |
+
continue
|
378 |
+
if quality_check({"augmentation": refined_candidate}, config):
|
379 |
+
full_candidate = {
|
380 |
+
"original_id": ex.id,
|
381 |
+
"augmentation": refined_candidate,
|
382 |
+
"strategy": strategy,
|
383 |
+
"metrics": metrics
|
384 |
+
}
|
385 |
+
augmented_candidates.append(full_candidate)
|
386 |
+
logger.info(f"Accepted candidate for {ex.id} (Total accepted: {len(augmented_candidates)})")
|
387 |
+
if len(augmented_candidates) >= target_count:
|
388 |
+
break
|
389 |
+
except Exception as e:
|
390 |
+
logger.error(f"Error generating augmentation for {ex.id}: {e}")
|
391 |
+
attempts += 1
|
392 |
+
if len(augmented_candidates) < target_count:
|
393 |
+
logger.warning(f"Only {len(augmented_candidates)} candidates generated after {attempts} attempts.")
|
394 |
+
return augmented_candidates
|
395 |
+
|
396 |
+
def deduplicate_augmentations(augmentations: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
|
397 |
+
"""
|
398 |
+
Remove duplicate augmentations based on hashable keys.
|
399 |
+
"""
|
400 |
+
seen = set()
|
401 |
+
unique_aug = []
|
402 |
+
for aug in augmentations:
|
403 |
+
key = (make_hashable(aug["augmentation"].get("augmented_input")),
|
404 |
+
make_hashable(aug["augmentation"].get("augmented_output")))
|
405 |
+
if key not in seen:
|
406 |
+
seen.add(key)
|
407 |
+
unique_aug.append(aug)
|
408 |
+
logger.info(f"Deduplicated to {len(unique_aug)} unique augmentations.")
|
409 |
+
return unique_aug
|
410 |
+
|
411 |
+
def format_for_openai(augmentations: List[Dict[str, Any]], system_message: str) -> str:
|
412 |
+
"""
|
413 |
+
Format augmentations in OpenAI fine-tuning JSONL format.
|
414 |
+
"""
|
415 |
+
output_lines = []
|
416 |
+
sys_msg = system_message.strip() if system_message and system_message.strip() else ""
|
417 |
+
for aug in augmentations:
|
418 |
+
user_val = flatten_content(fix_content(get_text(aug["augmentation"].get("augmented_input", "")).strip()))
|
419 |
+
assistant_val = flatten_content(fix_content(get_text(aug["augmentation"].get("augmented_output", "")).strip()))
|
420 |
+
record = {
|
421 |
+
"messages": [
|
422 |
+
{"role": "system", "content": sys_msg},
|
423 |
+
{"role": "user", "content": user_val},
|
424 |
+
{"role": "assistant", "content": assistant_val}
|
425 |
+
]
|
426 |
+
}
|
427 |
+
if validate_jsonl_record(record):
|
428 |
+
output_lines.append(json.dumps(record))
|
429 |
+
else:
|
430 |
+
logger.error(f"Record validation failed: {record}")
|
431 |
+
logger.info(f"Formatted {len(output_lines)} records in OpenAI fine-tuning format.")
|
432 |
+
return "\n".join(output_lines)
|
433 |
+
|
434 |
+
def format_for_gemini(augmentations: List[Dict[str, Any]]) -> str:
|
435 |
+
"""
|
436 |
+
Format augmentations in Gemini fine-tuning JSONL format.
|
437 |
+
"""
|
438 |
+
output_lines = []
|
439 |
+
for aug in augmentations:
|
440 |
+
user_val = flatten_content(fix_content(get_text(aug["augmentation"].get("augmented_input", "")).strip()))
|
441 |
+
assistant_val = flatten_content(fix_content(get_text(aug["augmentation"].get("augmented_output", "")).strip()))
|
442 |
+
record = {
|
443 |
+
"contents": [
|
444 |
+
{"role": "user", "parts": [{"text": user_val}]},
|
445 |
+
{"role": "model", "parts": [{"text": assistant_val}]}
|
446 |
+
]
|
447 |
+
}
|
448 |
+
if user_val and assistant_val:
|
449 |
+
output_lines.append(json.dumps(record))
|
450 |
+
else:
|
451 |
+
logger.error(f"Gemini record validation failed: {record}")
|
452 |
+
logger.info(f"Formatted {len(output_lines)} records in Gemini fine-tuning format.")
|
453 |
+
return "\n".join(output_lines)
|
454 |
+
|
455 |
+
def format_for_common(augmentations: List[Dict[str, Any]]) -> str:
|
456 |
+
"""
|
457 |
+
Format augmentations in a common JSONL format for both Mistral and LLama.
|
458 |
+
"""
|
459 |
+
output_lines = []
|
460 |
+
for aug in augmentations:
|
461 |
+
user_val = flatten_content(fix_content(get_text(aug["augmentation"].get("augmented_input", "")).strip()))
|
462 |
+
assistant_val = flatten_content(fix_content(get_text(aug["augmentation"].get("augmented_output", "")).strip()))
|
463 |
+
record = {
|
464 |
+
"messages": [
|
465 |
+
{"role": "user", "content": user_val},
|
466 |
+
{"role": "assistant", "content": assistant_val}
|
467 |
+
]
|
468 |
+
}
|
469 |
+
if user_val and assistant_val:
|
470 |
+
output_lines.append(json.dumps(record))
|
471 |
+
else:
|
472 |
+
logger.error(f"Common format record validation failed: {record}")
|
473 |
+
logger.info(f"Formatted {len(output_lines)} records in common JSONL format for Mistral/LLama.")
|
474 |
+
return "\n".join(output_lines)
|
475 |
+
|
476 |
+
def format_for_mistral(augmentations: List[Dict[str, Any]]) -> str:
|
477 |
+
"""
|
478 |
+
Format augmentations in Mistral fine-tuning JSONL format.
|
479 |
+
Uses the common format.
|
480 |
+
"""
|
481 |
+
return format_for_common(augmentations)
|
482 |
+
|
483 |
+
def format_for_llama(augmentations: List[Dict[str, Any]]) -> str:
|
484 |
+
"""
|
485 |
+
Format augmentations in LLama fine-tuning JSONL format.
|
486 |
+
Uses the common format.
|
487 |
+
"""
|
488 |
+
return format_for_common(augmentations)
|
489 |
+
|
490 |
+
# -----------------------------
|
491 |
+
# Optional: Load Existing Examples from JSONL
|
492 |
+
# -----------------------------
|
493 |
+
def load_examples_from_file(file_path: str, format_type: str = "openai") -> List[AugmentationExample]:
|
494 |
+
"""
|
495 |
+
Load input/output pairs from a JSONL file.
|
496 |
+
"""
|
497 |
+
examples = []
|
498 |
+
try:
|
499 |
+
with open(file_path, "r") as f:
|
500 |
+
for line in f:
|
501 |
+
line = line.strip()
|
502 |
+
if not line:
|
503 |
+
continue
|
504 |
+
record = json.loads(line)
|
505 |
+
if format_type.lower() == "openai":
|
506 |
+
msgs = record.get("messages", [])
|
507 |
+
if len(msgs) == 3:
|
508 |
+
user_text = msgs[1].get("content", "").strip()
|
509 |
+
assistant_text = msgs[2].get("content", "").strip()
|
510 |
+
if user_text and assistant_text:
|
511 |
+
examples.append(AugmentationExample(input_text=user_text, output_text=assistant_text))
|
512 |
+
elif format_type.lower() == "gemini":
|
513 |
+
contents = record.get("contents", [])
|
514 |
+
if len(contents) >= 2:
|
515 |
+
user_parts = contents[0].get("parts", [])
|
516 |
+
model_parts = contents[1].get("parts", [])
|
517 |
+
user_text = get_text(user_parts[0]) if user_parts else ""
|
518 |
+
assistant_text = get_text(model_parts[0]) if model_parts else ""
|
519 |
+
if user_text and assistant_text:
|
520 |
+
examples.append(AugmentationExample(input_text=user_text, output_text=assistant_text))
|
521 |
+
except Exception as e:
|
522 |
+
logger.error(f"Error loading examples from file: {e}")
|
523 |
+
logger.info(f"Loaded {len(examples)} examples from {file_path}")
|
524 |
+
return examples
|
525 |
+
|
526 |
+
# -----------------------------
|
527 |
+
# Pipeline Class
|
528 |
+
# -----------------------------
|
529 |
+
class FinetuningDataAugmentor:
|
530 |
+
"""
|
531 |
+
Encapsulates the entire augmentation pipeline.
|
532 |
+
"""
|
533 |
+
def __init__(self, config: AugmentationConfig):
|
534 |
+
self.config = config
|
535 |
+
self.normalized_examples = normalize_examples(config.examples)
|
536 |
+
self.strategy = determine_augmentation_strategy(config)
|
537 |
+
self.augmentations = []
|
538 |
+
|
539 |
+
def run_augmentation(self, target_count: int = 50) -> List[Dict[str, Any]]:
|
540 |
+
"""
|
541 |
+
Generate candidate augmentations, deduplicate, and store results.
|
542 |
+
"""
|
543 |
+
logger.info("Starting augmentation generation via LangChain Groq...")
|
544 |
+
candidates = generate_augmentations(self.normalized_examples, self.config, self.strategy, target_count=target_count)
|
545 |
+
logger.info(f"Generated {len(candidates)} candidate augmentations before deduplication.")
|
546 |
+
unique_candidates = deduplicate_augmentations(candidates)
|
547 |
+
logger.info(f"{len(unique_candidates)} unique augmentations after deduplication.")
|
548 |
+
self.augmentations = unique_candidates
|
549 |
+
return unique_candidates
|
550 |
+
|
551 |
+
def get_formatted_output(self, format_type: str = "openai") -> str:
|
552 |
+
"""
|
553 |
+
Return the final augmented data in the desired finetuning format.
|
554 |
+
"""
|
555 |
+
fmt = format_type.lower()
|
556 |
+
if fmt == "openai":
|
557 |
+
return format_for_openai(self.augmentations, self.config.system_message)
|
558 |
+
elif fmt == "gemini":
|
559 |
+
return format_for_gemini(self.augmentations)
|
560 |
+
elif fmt == "mistral":
|
561 |
+
return format_for_mistral(self.augmentations)
|
562 |
+
elif fmt == "llama":
|
563 |
+
return format_for_llama(self.augmentations)
|
564 |
+
else:
|
565 |
+
logger.error(f"Unknown format type: {format_type}. Defaulting to OpenAI format.")
|
566 |
+
return format_for_openai(self.augmentations, self.config.system_message)
|
567 |
+
|
568 |
+
def save_to_file(self, filename: str = "train.jsonl") -> None:
|
569 |
+
"""
|
570 |
+
Save the formatted augmented data to a file.
|
571 |
+
"""
|
572 |
+
output = self.get_formatted_output()
|
573 |
+
with open(filename, "w") as f:
|
574 |
+
f.write(output)
|
575 |
+
logger.info(f"Final augmented data saved to {filename}")
|
576 |
+
|
577 |
+
def run_review_interface(self) -> None:
|
578 |
+
"""
|
579 |
+
Launch the interactive review interface.
|
580 |
+
"""
|
581 |
+
from streamlit import runtime
|
582 |
+
formatted_data = self.get_formatted_output()
|
583 |
+
launch_review_app(formatted_data)
|
requirements.txt
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# List of dependencies
|
2 |
+
streamlit==1.42.2
|
3 |
+
langchain-groq==0.2.4
|
4 |
+
langchain-core==0.3.37
|
5 |
+
streamlit-ace==0.1.1
|
6 |
+
dotenv
|