Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

steps towards ravel in tlens #29

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

steps towards ravel in tlens #29

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
214 changes: 214 additions & 0 deletions evals/ravel/ravel/clean_prototype.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,214 @@
{
"cells": [
{
"cell_type": "code",
"execution_count": 1,
"metadata": {},
"outputs": [],
"source": [
"%load_ext autoreload\n",
"%autoreload 2\n",
"\n",
"import os\n",
"\n",
"REPO_DIR = f'{os.getcwd()}'\n",
"SRC_DIR = os.path.join(REPO_DIR, 'src')\n",
"MODEL_DIR = os.path.join(REPO_DIR, 'models')\n",
"DATA_DIR = os.path.join(REPO_DIR, 'data')\n",
"\n",
"for d in [MODEL_DIR, DATA_DIR]:\n",
" if not os.path.exists(d):\n",
" os.makedirs(d)\n",
"\n",
"\n",
"import sys\n",
"sys.path.append(REPO_DIR)\n",
"sys.path.append(SRC_DIR)\n",
"\n",
"import numpy as np\n",
"import random\n",
"import torch\n",
"import accelerate\n",
"# from nnsight import NNsight\n",
"from transformers import AutoModelForCausalLM, AutoTokenizer\n",
"from transformer_lens import HookedTransformer\n",
"\n",
"def set_seed(seed):\n",
" random.seed(seed)\n",
" np.random.seed(seed)\n",
" torch.manual_seed(seed)\n",
" torch.cuda.manual_seed_all(seed)\n",
"\n",
"set_seed(0)\n",
"\n",
"device = \"cpu\"\n",
"if torch.backends.mps.is_available():\n",
" device = \"mps\"\n",
"elif torch.cuda.is_available():\n",
" device = \"cuda\""
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Model"
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {},
"outputs": [
{
"ename": "FileNotFoundError",
"evalue": "[Errno 2] No such file or directory: '../../auth/hf_token.txt'",
"output_type": "error",
"traceback": [
"\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
"\u001b[0;31mFileNotFoundError\u001b[0m Traceback (most recent call last)",
"Cell \u001b[0;32mIn[3], line 4\u001b[0m\n\u001b[1;32m 1\u001b[0m \u001b[38;5;66;03m# Load model\u001b[39;00m\n\u001b[0;32m----> 4\u001b[0m \u001b[38;5;28;01mwith\u001b[39;00m \u001b[38;5;28;43mopen\u001b[39;49m\u001b[43m(\u001b[49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43m../../auth/hf_token.txt\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[38;5;124;43mr\u001b[39;49m\u001b[38;5;124;43m'\u001b[39;49m\u001b[43m)\u001b[49m \u001b[38;5;28;01mas\u001b[39;00m f:\n\u001b[1;32m 5\u001b[0m hf_token \u001b[38;5;241m=\u001b[39m f\u001b[38;5;241m.\u001b[39mread()\u001b[38;5;241m.\u001b[39mstrip()\n\u001b[1;32m 7\u001b[0m model_id \u001b[38;5;241m=\u001b[39m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mgoogle/gemma-2-2b\u001b[39m\u001b[38;5;124m\"\u001b[39m\n",
"File \u001b[0;32m~/miniconda3/envs/sae_eval/lib/python3.10/site-packages/IPython/core/interactiveshell.py:324\u001b[0m, in \u001b[0;36m_modified_open\u001b[0;34m(file, *args, **kwargs)\u001b[0m\n\u001b[1;32m 317\u001b[0m \u001b[38;5;28;01mif\u001b[39;00m file \u001b[38;5;129;01min\u001b[39;00m {\u001b[38;5;241m0\u001b[39m, \u001b[38;5;241m1\u001b[39m, \u001b[38;5;241m2\u001b[39m}:\n\u001b[1;32m 318\u001b[0m \u001b[38;5;28;01mraise\u001b[39;00m \u001b[38;5;167;01mValueError\u001b[39;00m(\n\u001b[1;32m 319\u001b[0m \u001b[38;5;124mf\u001b[39m\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mIPython won\u001b[39m\u001b[38;5;124m'\u001b[39m\u001b[38;5;124mt let you open fd=\u001b[39m\u001b[38;5;132;01m{\u001b[39;00mfile\u001b[38;5;132;01m}\u001b[39;00m\u001b[38;5;124m by default \u001b[39m\u001b[38;5;124m\"\u001b[39m\n\u001b[1;32m 320\u001b[0m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mas it is likely to crash IPython. If you know what you are doing, \u001b[39m\u001b[38;5;124m\"\u001b[39m\n\u001b[1;32m 321\u001b[0m \u001b[38;5;124m\"\u001b[39m\u001b[38;5;124myou can use builtins\u001b[39m\u001b[38;5;124m'\u001b[39m\u001b[38;5;124m open.\u001b[39m\u001b[38;5;124m\"\u001b[39m\n\u001b[1;32m 322\u001b[0m )\n\u001b[0;32m--> 324\u001b[0m \u001b[38;5;28;01mreturn\u001b[39;00m \u001b[43mio_open\u001b[49m\u001b[43m(\u001b[49m\u001b[43mfile\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[38;5;241;43m*\u001b[39;49m\u001b[43margs\u001b[49m\u001b[43m,\u001b[49m\u001b[43m \u001b[49m\u001b[38;5;241;43m*\u001b[39;49m\u001b[38;5;241;43m*\u001b[39;49m\u001b[43mkwargs\u001b[49m\u001b[43m)\u001b[49m\n",
"\u001b[0;31mFileNotFoundError\u001b[0m: [Errno 2] No such file or directory: '../../auth/hf_token.txt'"
]
}
],
"source": [
"# Load model\n",
"\n",
"\n",
"with open('../../auth/hf_token.txt', 'r') as f:\n",
" hf_token = f.read().strip()\n",
"\n",
"model_id = \"google/gemma-2-2b\"\n",
"model_name = \"gemma-2-2b\"\n",
"\n",
"torch.set_grad_enabled(False) # avoid blowing up mem\n",
"hf_model = AutoModelForCausalLM.from_pretrained(\n",
" model_id,\n",
" cache_dir=MODEL_DIR,\n",
" token=hf_token,\n",
" device_map=device,\n",
" low_cpu_mem_usage=True,\n",
" attn_implementation=\"eager\"\n",
")\n",
"\n",
"tokenizer = AutoTokenizer.from_pretrained(\n",
" model_id,\n",
" cache_dir=MODEL_DIR,\n",
" token=hf_token,\n",
")\n",
"tokenizer.pad_token = tokenizer.eos_token\n",
"tokenizer.padding_side = 'left'\n",
"VOCAB = sorted(tokenizer.vocab, key=tokenizer.vocab.get)\n",
"\n",
"layer_idx = 10\n",
"\n",
"\n",
"nnsight_model = NNsight(hf_model)\n",
"nnsight_tracer_kwargs = {'scan': True, 'validate': False, 'use_cache': False, 'output_attentions': False}"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Dataset Generation"
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {},
"outputs": [
{
"ename": "NameError",
"evalue": "name 'tokenizer' is not defined",
"output_type": "error",
"traceback": [
"\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
"\u001b[0;31mNameError\u001b[0m Traceback (most recent call last)",
"Cell \u001b[0;32mIn[2], line 3\u001b[0m\n\u001b[1;32m 1\u001b[0m \u001b[38;5;28;01mfrom\u001b[39;00m \u001b[38;5;21;01mravel_dataset_builder\u001b[39;00m \u001b[38;5;28;01mimport\u001b[39;00m RAVELEntityPromptData\n\u001b[0;32m----> 3\u001b[0m full_entity_dataset \u001b[38;5;241m=\u001b[39m RAVELEntityPromptData\u001b[38;5;241m.\u001b[39mfrom_files(\u001b[38;5;124m'\u001b[39m\u001b[38;5;124mcity\u001b[39m\u001b[38;5;124m'\u001b[39m, \u001b[38;5;124m'\u001b[39m\u001b[38;5;124mdata\u001b[39m\u001b[38;5;124m'\u001b[39m, \u001b[43mtokenizer\u001b[49m)\n\u001b[1;32m 4\u001b[0m \u001b[38;5;28mlen\u001b[39m(full_entity_dataset)\n",
"\u001b[0;31mNameError\u001b[0m: name 'tokenizer' is not defined"
]
}
],
"source": [
"from ravel_dataset_builder import RAVELEntityPromptData\n",
"\n",
"full_entity_dataset = RAVELEntityPromptData.from_files('city', 'data', tokenizer)\n",
"len(full_entity_dataset)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"sampled_entity_dataset = full_entity_dataset.downsample(1000)\n",
"print(f\"Number of prompts remaining: {len(sampled_entity_dataset)}\")\n",
"\n",
"prompt_max_length = 48\n",
"sampled_entity_dataset.generate_completions(nnsight_model, tokenizer, max_length=prompt_max_length+8, prompt_max_length=prompt_max_length)\n",
"\n",
"sampled_entity_dataset.evaluate_correctness()\n",
"\n",
"# Filter correct completions\n",
"correct_data = sampled_entity_dataset.filter_correct()\n",
"\n",
"# Filter top entities and templates\n",
"filtered_data = correct_data.filter_top_entities_and_templates(top_n_entities=400, top_n_templates_per_attribute=12)\n",
"\n",
"# Calculate average accuracy\n",
"accuracy = sampled_entity_dataset.calculate_average_accuracy()\n",
"print(f\"Average accuracy: {accuracy:.2%}\")\n",
"print(f\"Number of prompts remaining: {len(correct_data)}\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"correct_data.add_wikipedia_prompts('city', 'data', tokenizer, nnsight_model)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Experimental Interventions"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "base",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.10.14"
}
},
"nbformat": 4,
"nbformat_minor": 2
}
7 changes: 7 additions & 0 deletions evals/ravel/ravel/common_imports.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
from transformer_lens import HookedTransformer
import torch
import numpy as np
import pandas as pd
from torch import Tensor
import sys
from typing import List, Tuple, Dict, Any, Union, Literal, Optional
Binary file added evals/ravel/ravel/data.tgz
View file
Open in desktop
Binary file not shown.
Loading