File size: 3,611 Bytes
45424d1 095bd48 45424d1 095bd48 45424d1 095bd48 45424d1 095bd48 45424d1 095bd48 45424d1 095bd48 45424d1 095bd48 45424d1 |
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 |
# app.py (Hugging Face Space - runs GPT-2 + TransformerLens locally)
import os
import json
import streamlit as st
import requests
from datetime import datetime
# Import your model utilities (make sure model_utils.py is in the same repo on HF)
from model_utils import generate_text, run_activation_patching # uploaded at /mnt/data/model_utils.py
# Configure Render URL (set your actual Render URL here)
# Example: https://activation-patching-api.onrender.com
RENDER_API_BASE = st.secrets.get("render_url") or st.sidebar.text_input(
"Render API base URL", value="https://activation-patching-api.onrender.com"
)
SAVE_ENDPOINT = RENDER_API_BASE.rstrip("/") + "/save"
st.title("Mechanistic Analysis Interface (HF Space)")
st.write("This Streamlit app runs GPT-2 + activation patching locally, then saves metadata to your Render backend.")
prompt = st.text_area("Enter your sentence / prompt", height=150)
max_length = st.sidebar.slider("Max generation length", 20, 200, 60)
if st.button("Run Experiment"):
if not prompt.strip():
st.warning("Please enter a prompt.")
else:
with st.spinner("Generating text with GPT-2..."):
try:
generated = generate_text(prompt, max_length=max_length)
except Exception as e:
st.error(f"Error running generation: {e}")
raise
st.subheader("Generated Text")
st.write(generated)
with st.spinner("Running activation patching (TransformerLens)..."):
try:
activations = run_activation_patching(prompt)
except Exception as e:
st.error(f"Error running activation patching: {e}")
raise
st.subheader("Activation traces (sample)")
# activations can be large — summarize for display
sample = {k: (v.shape if hasattr(v, "shape") else type(v).__name__) for k, v in list(activations.items())[:10]}
st.json(sample)
# Ask explanation agent or placeholder — for now, simple summary:
explanation = "Explanation placeholder: top influencing layers ... (expand with LangGraph later)"
st.subheader("Explanation")
st.write(explanation)
# Prepare data to save to Render
payload = {
"prompt": prompt,
"generated_text": generated,
# Convert activations to a compact serializable form: keep shapes and optionally min/max
"activation_traces": json.dumps({
k: {
"shape": getattr(v, "shape", None),
"min": float(v.min()) if hasattr(v, "min") else None,
"max": float(v.max()) if hasattr(v, "max") else None
} for k, v in activations.items()
}),
"explanation": explanation
}
# Save to Render
try:
res = requests.post(SAVE_ENDPOINT, json=payload, timeout=30)
st.write("Save status:", res.status_code)
st.write("Save response:", res.text)
if res.ok:
data = res.json()
st.success(f"Experiment saved with ID {data.get('id')}")
else:
st.error(f"Failed to save experiment: {res.text}")
except Exception as e:
st.error(f"Error saving to Render: {e}")
st.markdown("---")
st.write("Notes:")
st.write("- This app runs heavy ML locally in the HF Space container; Render is used only to persist metadata.")
st.write("- If you want LangGraph explanation, we can call a low-cost open model here or run agents locally.")
|