Delete app

app/agents.py

@@ -1,23 +0,0 @@
-# This is a conceptual setup; use LangGraph syntax
-from langgraph import Agent, Workflow
-
-# Experiment Agent: runs GPT-2 + activation patching
-class ExperimentAgent(Agent):
-    def run(self, prompt):
-        from model_utils import generate_text, run_activation_patching
-        text = generate_text(prompt)
-        activations = run_activation_patching(prompt)
-        return {"generated_text": text, "activations": activations}
-
-# Explanation Agent: converts traces to NL explanation
-class ExplanationAgent(Agent):
-    def run(self, activations):
-        # placeholder: summarize important layers
-        explanation = "Layer 5 and 7 had the most influence on next token"
-        return explanation
-
-# Workflow connecting agents
-workflow = Workflow()
-workflow.add_agent("experiment", ExperimentAgent())
-workflow.add_agent("explanation", ExplanationAgent())
-workflow.connect("experiment", "explanation", lambda result: result["activations"])

app/api.py

@@ -1,33 +0,0 @@
-from fastapi import FastAPI
-from pydantic import BaseModel
-from model_utils import generate_text, run_activation_patching
-from db import init_db, save_experiment, get_experiment
-
-app = FastAPI()
-init_db()
-
-class ExperimentInput(BaseModel):
-    prompt: str
-
-@app.post("/generate")
-def generate(input: ExperimentInput):
-    text = generate_text(input.prompt)
-    activations = run_activation_patching(input.prompt)
-    # Placeholder for explanation (use LangGraph later)
-    explanation = "Explanation will be generated by agent"
-    exp_id = save_experiment(input.prompt, text, str(activations), explanation)
-    return {"id": exp_id, "generated_text": text, "activations": activations, "explanation": explanation}
-
-@app.get("/results/{id}")
-def get_results(id: int):
-    row = get_experiment(id)
-    if row:
-        return {
-            "id": row[0],
-            "prompt": row[1],
-            "generated_text": row[2],
-            "activation_traces": row[3],
-            "explanation": row[4],
-            "timestamp": row[5]
-        }
-    return {"error": "Experiment not found"}

app/app.py

@@ -1,25 +0,0 @@
-import streamlit as st
-import requests
-
-API_URL = "http://127.0.0.1:8000"
-
-st.title("Mechanistic Analysis Interface")
-
-prompt = st.text_area("Enter your sentence:")
-
-if st.button("Run Experiment"):
-    if prompt:
-        with st.spinner("Running GPT-2 + Activation Patching..."):
-            response = requests.post(f"{API_URL}/generate", json={"prompt": prompt})
-            data = response.json()
-        
-        st.subheader("Generated Text")
-        st.write(data["generated_text"])
-        
-        st.subheader("Activation Patching Traces")
-        st.write(data["activations"])
-        
-        st.subheader("Explanation")
-        st.write(data["explanation"])
-        
-        st.success(f"Experiment saved with ID: {data['id']}")

app/db.py

@@ -1,40 +0,0 @@
-import sqlite3
-from datetime import datetime
-
-DB_PATH = "results.db"
-
-def init_db():
-    conn = sqlite3.connect(DB_PATH)
-    c = conn.cursor()
-    c.execute("""
-        CREATE TABLE IF NOT EXISTS experiments (
-            id INTEGER PRIMARY KEY AUTOINCREMENT,
-            prompt TEXT,
-            generated_text TEXT,
-            activation_traces TEXT,
-            explanation TEXT,
-            timestamp TEXT
-        )
-    """)
-    conn.commit()
-    conn.close()
-
-def save_experiment(prompt, generated_text, activation_traces, explanation):
-    conn = sqlite3.connect(DB_PATH)
-    c = conn.cursor()
-    c.execute("""
-        INSERT INTO experiments (prompt, generated_text, activation_traces, explanation, timestamp)
-        VALUES (?, ?, ?, ?, ?)
-    """, (prompt, generated_text, activation_traces, explanation, datetime.now().isoformat()))
-    conn.commit()
-    experiment_id = c.lastrowid
-    conn.close()
-    return experiment_id
-
-def get_experiment(exp_id):
-    conn = sqlite3.connect(DB_PATH)
-    c = conn.cursor()
-    c.execute("SELECT * FROM experiments WHERE id=?", (exp_id,))
-    row = c.fetchone()
-    conn.close()
-    return row

app/model_utils.py

@@ -1,36 +0,0 @@
-from transformers import GPT2LMHeadModel, GPT2Tokenizer
-import torch
-from transformer_lens import HookedTransformer
-
-def load_gpt2():
-    model = GPT2LMHeadModel.from_pretrained("gpt2")
-    tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
-    return model, tokenizer
-
-def generate_text(prompt, max_length=50):
-    model, tokenizer = load_gpt2()
-    inputs = tokenizer(prompt, return_tensors="pt")
-    output = model.generate(**inputs, max_length=max_length)
-    text = tokenizer.decode(output[0], skip_special_tokens=True)
-    return text
-
-# Example: HookedTransformer for activation patching
-def run_activation_patching(prompt):
-    model = HookedTransformer.from_pretrained("gpt2-small")
-    tokens = model.to_tokens(prompt)
-    
-    # Example: capture activations per layer
-    activations = {}
-    def hook_fn_hook(value, hook):
-        activations[hook.name] = value.detach().cpu().numpy()
-    
-    hooks = []
-    for i in range(model.cfg.n_layers):
-        hooks.append(model.hook.add_hook(f"blocks.{i}.mlp.hook_post", hook_fn_hook))
-    
-    logits = model(tokens)
-    # Remove hooks
-    for h in hooks:
-        h.remove()
-    
-    return activations