Update app.py
Browse files
app.py
CHANGED
|
@@ -1,23 +1,13 @@
|
|
| 1 |
-
#!/usr/bin/env python
|
| 2 |
-
"""HFΒ Space for the *Marinβ8BβInstruct* research preview
|
| 3 |
-
-----------------------------------------------------
|
| 4 |
-
A lightweight Gradio interface that
|
| 5 |
-
β’ streams chat completions from the `marin-community/marin-8b-instruct` model
|
| 6 |
-
β’ lets testers submit structured feedback (UX ratingΒ + freeβtext)
|
| 7 |
-
β’ appends feedback to a local JSONL *and* merges it into a private Hub dataset
|
| 8 |
-
The dataset is never overwritten: we always pull, merge, deduplicate, and push.
|
| 9 |
-
"""
|
| 10 |
-
|
| 11 |
from __future__ import annotations
|
| 12 |
|
| 13 |
-
#
|
| 14 |
import json
|
| 15 |
import os
|
| 16 |
import time
|
| 17 |
import uuid
|
| 18 |
from threading import Thread
|
| 19 |
|
| 20 |
-
#
|
| 21 |
import gradio as gr
|
| 22 |
from gradio_modal import Modal
|
| 23 |
from transformers import (
|
|
@@ -29,7 +19,7 @@ from datasets import Dataset, load_dataset, concatenate_datasets, DownloadMode
|
|
| 29 |
from huggingface_hub import HfApi, login
|
| 30 |
import spaces
|
| 31 |
|
| 32 |
-
# ββββββββββββββββββββββββββββ model
|
| 33 |
checkpoint = "marin-community/marin-8b-instruct"
|
| 34 |
device = "cuda" # the Space runner gives us a GPU
|
| 35 |
|
|
@@ -37,8 +27,8 @@ device = "cuda" # the Space runner gives us a GPU
|
|
| 37 |
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
|
| 38 |
model = AutoModelForCausalLM.from_pretrained(checkpoint).to(device)
|
| 39 |
|
| 40 |
-
# feedback
|
| 41 |
-
DATASET_REPO = "WillHeld/model-feedback" #
|
| 42 |
DATA_DIR = "./feedback_data"
|
| 43 |
DATA_FILE = "feedback.jsonl"
|
| 44 |
os.makedirs(DATA_DIR, exist_ok=True)
|
|
@@ -57,7 +47,7 @@ def save_feedback_locally(conversation: list[dict[str, str]],
|
|
| 57 |
"feedback": feedback_text,
|
| 58 |
}
|
| 59 |
fp = os.path.join(DATA_DIR, DATA_FILE)
|
| 60 |
-
with open(fp, "a", encoding="utf
|
| 61 |
f.write(json.dumps(record, ensure_ascii=False) + "\n")
|
| 62 |
return record["id"]
|
| 63 |
|
|
@@ -70,7 +60,7 @@ def push_feedback_to_hub(hf_token: str | None = None) -> bool: # noqa: C901
|
|
| 70 |
1. Authenticate with `hf_token` (fall back to $HF_TOKEN env).
|
| 71 |
2. Load *local* feedback just written in `feedback.jsonl`.
|
| 72 |
3. Pull existing remote split (if any); concat & `unique("id")`.
|
| 73 |
-
4. Push the merged dataset back.
|
| 74 |
"""
|
| 75 |
|
| 76 |
hf_token = hf_token or os.getenv("HF_TOKEN")
|
|
@@ -85,7 +75,7 @@ def push_feedback_to_hub(hf_token: str | None = None) -> bool: # noqa: C901
|
|
| 85 |
return False
|
| 86 |
|
| 87 |
# local rows β Dataset
|
| 88 |
-
with open(fp, encoding="utf
|
| 89 |
local_ds = Dataset.from_list([json.loads(l) for l in f])
|
| 90 |
|
| 91 |
# try to pull remote
|
|
@@ -153,7 +143,7 @@ def generate_response(message: str,
|
|
| 153 |
partial = ""
|
| 154 |
for token in streamer:
|
| 155 |
partial += token
|
| 156 |
-
yield partial, history # 1st
|
| 157 |
|
| 158 |
# once finished, commit assistant reply to history
|
| 159 |
history.append({"role": "assistant", "content": partial})
|
|
@@ -168,12 +158,12 @@ def submit_feedback(conversation_state: list[dict[str, str]],
|
|
| 168 |
save_feedback_locally(conversation_state, satisfaction, feedback_text)
|
| 169 |
pushed = push_feedback_to_hub()
|
| 170 |
if pushed:
|
| 171 |
-
return "β
Thanks!
|
| 172 |
return "β οΈ Saved locally; Hub push failed. Check server logs."
|
| 173 |
|
| 174 |
# ββββββββββββββββββββββββββββ UI layout ββββββββββββββββββββββββββββββββββββ
|
| 175 |
|
| 176 |
-
with gr.Blocks(title="Marin
|
| 177 |
# state object to surface chat history to the feedback form
|
| 178 |
conversation_state = gr.State([])
|
| 179 |
|
|
@@ -186,7 +176,7 @@ with gr.Blocks(title="Marinβ8B Research Preview") as demo:
|
|
| 186 |
gr.Slider(0.1, 2.0, value=0.7, step=0.1,
|
| 187 |
label="Temperature"),
|
| 188 |
gr.Slider(0.1, 1.0, value=0.9, step=0.05,
|
| 189 |
-
label="Top
|
| 190 |
additional_outputs=[conversation_state],
|
| 191 |
type="messages",
|
| 192 |
)
|
|
@@ -198,7 +188,7 @@ with gr.Blocks(title="Marinβ8B Research Preview") as demo:
|
|
| 198 |
# feedback modal (hidden by default)
|
| 199 |
with Modal(visible=False) as fb_modal:
|
| 200 |
gr.Markdown("## Research Preview Feedback")
|
| 201 |
-
gr.Markdown("We appreciate your help improving Marin
|
| 202 |
|
| 203 |
sat_radio = gr.Radio([
|
| 204 |
"Very satisfied", "Satisfied", "Neutral",
|
|
@@ -211,8 +201,8 @@ with gr.Blocks(title="Marinβ8B Research Preview") as demo:
|
|
| 211 |
status_box = gr.Textbox(label="Status", interactive=False)
|
| 212 |
|
| 213 |
# interactions
|
| 214 |
-
|
| 215 |
-
|
| 216 |
|
| 217 |
send_btn.click(
|
| 218 |
submit_feedback,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
from __future__ import annotations
|
| 2 |
|
| 3 |
+
# -- standard lib
|
| 4 |
import json
|
| 5 |
import os
|
| 6 |
import time
|
| 7 |
import uuid
|
| 8 |
from threading import Thread
|
| 9 |
|
| 10 |
+
# -- third-party deps (declared in requirements.txt of the Space)
|
| 11 |
import gradio as gr
|
| 12 |
from gradio_modal import Modal
|
| 13 |
from transformers import (
|
|
|
|
| 19 |
from huggingface_hub import HfApi, login
|
| 20 |
import spaces
|
| 21 |
|
| 22 |
+
# ββββββββββββββββββββββββββββ model & constants βββββββββββββββββββββββββββββ
|
| 23 |
checkpoint = "marin-community/marin-8b-instruct"
|
| 24 |
device = "cuda" # the Space runner gives us a GPU
|
| 25 |
|
|
|
|
| 27 |
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
|
| 28 |
model = AutoModelForCausalLM.from_pretrained(checkpoint).to(device)
|
| 29 |
|
| 30 |
+
# feedback dataset details
|
| 31 |
+
DATASET_REPO = "WillHeld/model-feedback" # <-- change to your namespace if needed
|
| 32 |
DATA_DIR = "./feedback_data"
|
| 33 |
DATA_FILE = "feedback.jsonl"
|
| 34 |
os.makedirs(DATA_DIR, exist_ok=True)
|
|
|
|
| 47 |
"feedback": feedback_text,
|
| 48 |
}
|
| 49 |
fp = os.path.join(DATA_DIR, DATA_FILE)
|
| 50 |
+
with open(fp, "a", encoding="utf-8") as f:
|
| 51 |
f.write(json.dumps(record, ensure_ascii=False) + "\n")
|
| 52 |
return record["id"]
|
| 53 |
|
|
|
|
| 60 |
1. Authenticate with `hf_token` (fall back to $HF_TOKEN env).
|
| 61 |
2. Load *local* feedback just written in `feedback.jsonl`.
|
| 62 |
3. Pull existing remote split (if any); concat & `unique("id")`.
|
| 63 |
+
4. Push the merged dataset back. Never deletes remote shards β safe.
|
| 64 |
"""
|
| 65 |
|
| 66 |
hf_token = hf_token or os.getenv("HF_TOKEN")
|
|
|
|
| 75 |
return False
|
| 76 |
|
| 77 |
# local rows β Dataset
|
| 78 |
+
with open(fp, encoding="utf-8") as f:
|
| 79 |
local_ds = Dataset.from_list([json.loads(l) for l in f])
|
| 80 |
|
| 81 |
# try to pull remote
|
|
|
|
| 143 |
partial = ""
|
| 144 |
for token in streamer:
|
| 145 |
partial += token
|
| 146 |
+
yield partial, history # 1st out = msg, 2nd out = state
|
| 147 |
|
| 148 |
# once finished, commit assistant reply to history
|
| 149 |
history.append({"role": "assistant", "content": partial})
|
|
|
|
| 158 |
save_feedback_locally(conversation_state, satisfaction, feedback_text)
|
| 159 |
pushed = push_feedback_to_hub()
|
| 160 |
if pushed:
|
| 161 |
+
return "β
Thanks! Your feedback is safely stored."
|
| 162 |
return "β οΈ Saved locally; Hub push failed. Check server logs."
|
| 163 |
|
| 164 |
# ββββββββββββββββββββββββββββ UI layout ββββββββββββββββββββββββββββββββββββ
|
| 165 |
|
| 166 |
+
with gr.Blocks(title="Marin-8B Research Preview") as demo:
|
| 167 |
# state object to surface chat history to the feedback form
|
| 168 |
conversation_state = gr.State([])
|
| 169 |
|
|
|
|
| 176 |
gr.Slider(0.1, 2.0, value=0.7, step=0.1,
|
| 177 |
label="Temperature"),
|
| 178 |
gr.Slider(0.1, 1.0, value=0.9, step=0.05,
|
| 179 |
+
label="Top-P")],
|
| 180 |
additional_outputs=[conversation_state],
|
| 181 |
type="messages",
|
| 182 |
)
|
|
|
|
| 188 |
# feedback modal (hidden by default)
|
| 189 |
with Modal(visible=False) as fb_modal:
|
| 190 |
gr.Markdown("## Research Preview Feedback")
|
| 191 |
+
gr.Markdown("We appreciate your help improving Marin-8B! β¨")
|
| 192 |
|
| 193 |
sat_radio = gr.Radio([
|
| 194 |
"Very satisfied", "Satisfied", "Neutral",
|
|
|
|
| 201 |
status_box = gr.Textbox(label="Status", interactive=False)
|
| 202 |
|
| 203 |
# interactions
|
| 204 |
+
# open the modal without custom JS β use Modal update
|
| 205 |
+
report_btn.click(lambda: Modal(visible=True), None, fb_modal)
|
| 206 |
|
| 207 |
send_btn.click(
|
| 208 |
submit_feedback,
|