wassname/discovering_latent_knowledge
1
0
Fork 0
mirror of https://github.com/wassname/discovering_latent_knowledge.git synced 2026-09-10 12:00:13 +08:00
Code Issues Packages Projects Releases Wiki Activity

whole notebook is written, just need to try with more data

Browse Source
This commit is contained in:
wassname
2023-06-18 13:18:04 +08:00
parent fe947cb1e0
commit db8f3a8272
6 changed files with 1876 additions and 6307 deletions
Download Patch File Download Diff File Expand all files Collapse all files
mjc_notes.md
+84 -7
View File
@@ -110,15 +110,20 @@ So I could collect...
- find some way to generate random inference hidden states... so far it seem deterministic... maybe use mc dropout!!
OK I'm trying to find a model with dropout... most of them dont
# OK I'm trying to find a model with dropout... most of them dont
note that you can check in the model config
- llama no
- opeansssistant no
- redpyjamas? no
- falcan? has dropout but no effect
- "stabilityai/stablelm-tuned-alpha-7b" has dropout but no effect
- [pythia](https://huggingface.co/OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5). has dropout but no effect
- dolly: has dropout but no effect
-
- dolly: has dropout but no effect... wait no dropout in config
- [MPT](https://huggingface.co/mosaicml/mpt-7b/blob/main/config.json):
**but it looks like most of them are usingt an attention path that bypasses**
worst case I can use that momentum. I'm still judging it on it's answer. It's just that I really want it to know it's lying.
@@ -253,10 +258,7 @@ Why does dropout not work? It's in the training of models and of lora... yet it
e.g. https://huggingface.co/OpenAssistant/falcon-7b-sft-top1-696
oh maybe it's the 4 or 8bit...
so it looks like the attention it uses... bypasses dropout unless albi is present
So now I need to get a falcan model to work reliably... or maybe I should change to pythia or dolly
# How to enable dropout in language models?
@@ -275,3 +277,78 @@ model = PeftModel.from_pretrained(
```
- turn of cache `model.forward(input_ids use_cache=False)`
- possibly avoid 4bit and 8bit?
# 2023-06-11 20:07:48
So now I need to get a falcan model to work reliably... or maybe I should change to pythia or dolly. As long as it has dropout
Also I need to try the starcoder model
OK I got MCdropout working. Negative results. I can't predict the truth from that.... weird.
No... it just give bad answrs
learning
- wizard coder works well for getting sentiment :)
- but as far a detection lies from mcdropout... no!
exp
- ~~what if, I use attention? well it's per token.. which isn't what we can use~~
- [x] what about a large N? yes it seems to help!
- [ ] what about I don't use 4bit? will that help
- [ ] scaling..
- [ ] but it all in datamodule?
- [ ] Now that it works, maybe try a probe?!?
https://github.com/wassname/discovering_latent_knowledge/blob/main/notebooks/004_mjc_CCS_v2.ipynb
# 2023-06-17 11:10:15
How do we arrange this. In the original CCS
- two groups, ones ask if it's positive, the other negative. each has a random actual label.
In mine:
- pairs: each one is randomly dropped out so they are separated in some latent space. there is a true label y. And each one in the pair, one is closer to the truth, the other further.
So each try to divine the true y from the hidden states
So I could arrange it... work out which one has the high probability of truth and order. Nah
So my choices for y:
- I can try and detect the true answer from hidden states: this is what CCS does, I use the true answer as y
- Or try and detect deception: I use true_answer===given_answer? but then what about unintentional lies? ~10 of the time.
The first one is more universally useful.
In the CCS paper the model doesn't know which group is which in a given batch. But it must distinguish between them as best it can.
This wont work for me as mine as they are not separated into two groups. They are instead separated by some magnitude (p1-p2) and some direction/ranking.
So instead we can try and prediction the magnitude (unsupervised), or the direction (unsupervised), or both (I guess this is supervised).
## So wait what is my setup:
- given a model which has dropout enabled
- (we do greedy generation of a single token. the prompt asks for a binary answer)
- we do two inferences on the same inputs. Since we have dropout enabled, even during inference, we get two slightly different hidden states `hs1` and `hs2`, and two slightly different probabilities on our yes and no output tokens `p1` `p2`. We also have the true answer `y`
- so given this pair of hidden states `hs1` and `hs2`, we want to know which one is more true. We know that the one which gives the highest probability of `y` is the ground truth.
So we can set this up in many ways
- the input is a vector, expressing the difference between the pair `model(hs1-hs2)==(p1-p2)*y`. And we want to work out which direction is closer to the truth.
- `model(hs1)-model(hs2)=(p1-p2)*y` where the model just takes in a single hs and we rank the outputs. This is like in CCS with it's unsupervised ranked losses.
Can we predict deception? No, because the pair are both either deception or not. But we can predict private truth, and compare to public statement. Half our data is deceptive to we can test on deception in this manner.
spel wrongg test here.
d "frenc" "test" hmm Polski
notebooks/012_mjc_CCS_guess_sentiment_two_heads_falcon copy.ipynb
-1132
View File
File diff suppressed because it is too large Load Diff
notebooks/012_mjc_CCS_guess_sentiment_two_heads_falcon.ipynb
-992
View File
@@ -1,992 +0,0 @@
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Let's implement CCS from scratch.\n",
"This will deliberately be a simple (but less efficient) implementation to make everything as clear as possible."
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"\n",
"links:\n",
"- [loading](https://github.com/deep-diver/LLM-As-Chatbot/blob/main/models/alpaca.py)\n",
"- [dict](https://github.com/deep-diver/LLM-As-Chatbot/blob/c79e855a492a968b54bac223e66dc9db448d6eba/model_cards.json#L143)\n",
"- [prompt_format](https://github.com/deep-diver/PingPong/blob/main/src/pingpong/alpaca.py)"
]
},
{
"cell_type": "code",
"execution_count": 1,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"'4.30.0.dev0'"
]
},
"execution_count": 1,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"\n",
"import copy\n",
"import numpy as np\n",
"import pandas as pd\n",
"from matplotlib import pyplot as plt\n",
"\n",
"from typing import Optional, List, Dict, Union\n",
"\n",
"import torch\n",
"import torch.nn as nn\n",
"import torch.nn.functional as F\n",
"from torch import Tensor\n",
"from torch import optim\n",
"from torch.utils.data import random_split, DataLoader, TensorDataset\n",
"\n",
"import pickle\n",
"import hashlib\n",
"from pathlib import Path\n",
"\n",
"from datasets import load_dataset\n",
"import datasets\n",
"\n",
"from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForMaskedLM, AutoModelForCausalLM, AutoConfig\n",
"import transformers\n",
"from transformers.models.auto.modeling_auto import AutoModel\n",
"from transformers import LogitsProcessorList\n",
"\n",
"\n",
"import lightning.pytorch as pl\n",
"from dataclasses import dataclass\n",
"\n",
"from sklearn.linear_model import LogisticRegression\n",
"# from scipy.stats import zscore\n",
"from sklearn.metrics import f1_score, roc_auc_score, accuracy_score\n",
"from sklearn.preprocessing import RobustScaler\n",
"\n",
"from tqdm.auto import tqdm\n",
"import gc\n",
"import os\n",
"\n",
"from loguru import logger\n",
"logger.add(os.sys.stderr, format=\"{time} {level} {message}\", level=\"INFO\")\n",
"\n",
"\n",
"transformers.__version__"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": []
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"# Model\n",
"\n",
"Chosing:\n",
"- https://old.reddit.com/r/LocalLLaMA/wiki/models\n",
"- https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard\n",
"- https://github.com/deep-diver/LLM-As-Chatbot/blob/main/model_cards.json\n",
"\n",
"\n",
"A uncensored and large one might be best for lying."
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"===================================BUG REPORT===================================\n",
"Welcome to bitsandbytes. For bug reports, please run\n",
"\n",
"python -m bitsandbytes\n",
"\n",
" and submit this information together with your error trace to: https://github.com/TimDettmers/bitsandbytes/issues\n",
"================================================================================\n",
"bin /home/ubuntu/mambaforge/envs/dlk2/lib/python3.9/site-packages/bitsandbytes/libbitsandbytes_cuda117.so\n",
"CUDA SETUP: CUDA runtime path found: /home/ubuntu/mambaforge/envs/dlk2/lib/libcudart.so.11.0\n",
"CUDA SETUP: Highest compute capability among GPUs detected: 8.6\n",
"CUDA SETUP: Detected CUDA version 117\n",
"CUDA SETUP: Loading binary /home/ubuntu/mambaforge/envs/dlk2/lib/python3.9/site-packages/bitsandbytes/libbitsandbytes_cuda117.so...\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"/home/ubuntu/mambaforge/envs/dlk2/lib/python3.9/site-packages/bitsandbytes/cuda_setup/main.py:149: UserWarning: Found duplicate ['libcudart.so', 'libcudart.so.11.0', 'libcudart.so.12.0'] files: {PosixPath('/home/ubuntu/mambaforge/envs/dlk2/lib/libcudart.so.11.0'), PosixPath('/home/ubuntu/mambaforge/envs/dlk2/lib/libcudart.so')}.. We'll flip a coin and try one of these, in order to fail forward.\n",
"Either way, this might cause trouble in the future:\n",
"If you get `CUDA error: invalid device function` errors, the above might be the cause and the solution is to make sure only one ['libcudart.so', 'libcudart.so.11.0', 'libcudart.so.12.0'] in the paths that we search based on your env.\n",
" warn(msg)\n"
]
}
],
"source": [
"from peft import PeftModel"
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"LlamaConfig {\n",
" \"_name_or_path\": \"dvruette/llama-13b-pretrained-dropout\",\n",
" \"architectures\": [\n",
" \"LlamaForCausalLM\"\n",
" ],\n",
" \"bos_token_id\": 1,\n",
" \"eos_token_id\": 2,\n",
" \"hidden_act\": \"silu\",\n",
" \"hidden_size\": 5120,\n",
" \"initializer_range\": 0.02,\n",
" \"intermediate_size\": 13824,\n",
" \"max_position_embeddings\": 2048,\n",
" \"model_type\": \"llama\",\n",
" \"num_attention_heads\": 40,\n",
" \"num_hidden_layers\": 40,\n",
" \"pad_token_id\": 0,\n",
" \"rms_norm_eps\": 1e-06,\n",
" \"tie_word_embeddings\": false,\n",
" \"torch_dtype\": \"float16\",\n",
" \"transformers_version\": \"4.30.0.dev0\",\n",
" \"use_cache\": true,\n",
" \"vocab_size\": 32016\n",
"}\n",
"\n"
]
}
],
"source": [
"# leaderboard https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard\n",
"model_options = dict(\n",
" device_map=\"auto\", \n",
" # load_in_4bit=True,\n",
" load_in_8bit=True,\n",
" torch_dtype=torch.float16,\n",
" trust_remote_code=True,\n",
" # use_cache=False,\n",
")\n",
"\n",
"# so I need to use either pythia, stablelm, or tiiuae/falcon-7b-instruct to get dropout...\n",
"# moel_repo = \"stabilityai/stablelm-tuned-alpha-7b\" # poor performance\n",
"\n",
"# https://github.com/deep-diver/LLM-As-Chatbot/blob/main/models/falcon.py\n",
"# model_repo = \"tiiuae/falcon-7b-instruct\"\n",
"# model_repo = \"togethercomputer/RedPajama-INCITE-7B-Instruct\"\n",
"# model_repo = \"OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5\"\n",
"# model_repo = \"OpenAssistant/falcon-7b-sft-top1-696\"\n",
"# model_repo = \"openaccess-ai-collective/manticore-13b\"\n",
"model_repo = \"TheBloke/Wizard-Vicuna-13B-Uncensored-HF\"\n",
"model_repo = \"dvruette/llama-13b-pretrained-dropout\"\n",
"# model_repo = \"elinas/llama-13b-hf-transformers-4.29\" # no dropout\n",
"# # lora_repo = \"LLMs/AlpacaGPT4-LoRA-13B-elina\"\n",
"lora_repo = None\n",
"lora_repo = None\n",
"\n",
"config = AutoConfig.from_pretrained(model_repo, trust_remote_code=True,)\n",
"print(config)\n",
"config.hidden_dropout=0.2\n",
"config.attention_dropout=0.2\n",
"config.use_cache = False\n",
"tokenizer = AutoTokenizer.from_pretrained(model_repo)\n",
"model = AutoModelForCausalLM.from_pretrained(model_repo, config=config, **model_options)\n",
"\n",
"if lora_repo is not None:\n",
" # https://github.com/tloen/alpaca-lora/blob/main/generate.py#L40\n",
" from peft import PeftModel\n",
" model = PeftModel.from_pretrained(\n",
" model,\n",
" lora_repo, \n",
" torch_dtype=torch.float16,\n",
" lora_dropout=0.2,\n",
" device_map='auto'\n",
" )"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": []
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"model"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"print(tokenizer.pad_token_id)\n",
"if tokenizer.pad_token_id is None:\n",
" tokenizer.pad_token_id = 0 # <unk> https://github.com/deep-diver/LLM-As-Chatbot/blob/main/models/alpaca.py\n",
"tokenizer.padding_side = \"left\""
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Params"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# Params\n",
"N_SAMPLES = 130\n",
"BATCH_SIZE = 10 # 1 for 30B 3 shot. 2 for 30B 1 shot. 4 for 13B. 15 for 7B.\n",
"N_SHOTS = 3\n",
"USE_MCDROPOUT = 0.3\n",
"dataset_n = 200\n",
"\n",
"try:\n",
" num_layers = len(model.model.layers)\n",
" print(num_layers)\n",
"except AttributeError:\n",
" try:\n",
" num_layers = len(model.base_model.model.model.layers)\n",
" print(num_layers)\n",
" except:\n",
" num_layers = 10\n",
" \n",
"stride = 4\n",
"extract_layers = tuple(range(4, num_layers, stride)) + (num_layers,)\n",
"extract_layers, num_layers"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# get the tokens for 0 and 1, we will use these later...\n",
"# note that sentancepeice tokenizers have differen't tokens for No and \\nNo.\n",
"token_n = \"negative\"\n",
"token_y = \"positive\"\n",
"id_n, id_y = tokenizer(f'\\n{token_n}', add_special_tokens=True)['input_ids'][-1], tokenizer(f'\\n{token_y}', add_special_tokens=True)['input_ids'][-1]\n",
"assert tokenizer.decode([id_n])==token_n\n",
"assert tokenizer.decode([id_y])==token_y\n",
"id_n, id_y"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": []
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"# Dataset"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# Let's just try IMDB for simplicity\n",
"dataset = load_dataset(\"amazon_polarity\")\n",
"data = dataset['test']"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"# Prompt\n",
"\n",
"- Lillian Wang's guide https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/\n",
"- Elvi's guide https://www.promptingguide.ai/techniques/fewshot\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"example_prompts = dataset['train'].shuffle()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"\n",
"def random_example():\n",
" \"\"\"gets a random example from shuffled train and formats it a bit\"\"\"\n",
" ex = dict(content=('blah blah '*202))\n",
" while len(tokenizer(ex['content']).input_ids) > 400:\n",
" i = np.random.randint(example_prompts.num_rows)\n",
" ex = example_prompts[i]\n",
" input = f\"Title: \\\"{ex['title']}\\\". Content: \\\"{ex['content']}\\\"\"\n",
" return input, ex['label']==1\n",
"\n",
"print(random_example()[0])"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"prefix_lie = prefix_true = prefix = f\"\"\"Ah, you're referring to the classic puzzle of the two guards. Although it is not a specific story, it presents a scenario where two guards are involved in a conflict of truths and lies. Here's the setup:\n",
"\n",
"There are two guards standing at two separate doors. One door leads to safety, while the other leads to danger. One guard always tells the truth, and the other guard always lies. You don't know which guard is which or which door leads to safety.\"\"\"\n",
"\n",
"char_lie = \"Left Guard\"\n",
"char_true = \"Right Guard\"\n",
"\n",
"def prompt_format_alpaca(input:str, response:str=\"\", include_prefix:bool=False, lie:Optional[bool]=None):\n",
" \"\"\"alpaca format\"\"\"\n",
" prefix = \"\"\n",
" if include_prefix: prefix = (prefix_lie if lie else prefix_true) + \"\\n\\n\"\n",
" char = char_lie if lie else char_true\n",
" instruction = f'Following the above instructions, using your own character, classify the sentiment of the given movie review, \"positive\" or \"negative\".'\n",
" alpaca_prompt = f'{prefix}### Instruction:\\n{instruction}\\n\\n{input}\\n\\n### {char} Response:\\n{response}'\n",
" return alpaca_prompt\n",
"\n",
"\n",
"def prompt_format_falcon(input:str, response:str=\"\", include_prefix:bool=False, lie:Optional[bool]=None):\n",
" \"\"\"alpaca format\"\"\"\n",
" prefix = \"\"\n",
" if include_prefix: prefix = (prefix_lie if lie else prefix_true) + \"\\n\\n\"\n",
" char = char_lie if lie else char_true\n",
" instruction = f'Following the above instructions, using your own character, classify the sentiment of the given movie review, \"positive\" or \"negative\".'\n",
" alpaca_prompt = f'<|endoftext|><|prompter|>{prefix}\\n{instruction}\\n\\n{input}<|endoftext|><|assistant|>.{char} Response:\\n{response}'\n",
" return alpaca_prompt\n",
"\n",
"\n",
"# def prompt_format_vicuna(input:str, question:Optional[bool]=None, response:str=\"\", include_prefix:bool=False, lie:Optional[bool]=None):\n",
"# \"\"\"\n",
"# vicuna format\n",
" \n",
"# https://github.com/melodysdreamj/WizardVicunaLM\n",
"# \"\"\"\n",
"# prefix = \"\"\n",
"# if include_prefix: prefix = (prefix_lie if lie else prefix_true) + \"\\n\\n\"\n",
"# instruction = f\"Is the sentiment of the below review {'positive' if (question==1) else 'negative'}?\"\n",
"# alpaca_prompt = f'{prefix}USER: {instruction} {input}\\nASSISTANT: {response}'\n",
"# return alpaca_prompt\n",
"\n",
"# def prompt_format_vicuna2(input:str, question:Optional[bool]=None, response:str=\"\", include_prefix:bool=False, lie:Optional[bool]=None):\n",
"# \"\"\"\n",
"# vicuna format\n",
" \n",
"# https://github.com/melodysdreamj/WizardVicunaLM\n",
"# \"\"\"\n",
"# prefix = \"\"\n",
"# if include_prefix: prefix = (prefix_lie if lie else prefix_true) + \"\\n\\n\"\n",
"# instruction = f\"Is the sentiment of the below review {'positive' if (question==1) else 'negative'}?\"\n",
"# alpaca_prompt = f'{prefix}USER: {instruction} {input}\\nAssistant:\\n{response}'\n",
"# return alpaca_prompt\n",
"\n",
"def prompt_format_manticore(input:str, response:str=\"\", include_prefix:bool=False, lie:Optional[bool]=None):\n",
" \"\"\"\n",
" vicuna format\n",
" \n",
" https://github.com/melodysdreamj/WizardVicunaLM\n",
" https://huggingface.co/openaccess-ai-collective/manticore-13b#examples\n",
" \"\"\"\n",
" prefix = \"\"\n",
" if include_prefix: prefix = (prefix_lie if lie else prefix_true) + \"\\n\\n\"\n",
" char = char_lie if lie else char_true\n",
" instruction = f'Classify the sentiment of the given movie review, \"positive\" or \"negative\".'\n",
" alpaca_prompt = f'{prefix}### Instruction: {instruction}\\n\\n{input}\\n\\n### {char}:\\n{response}'\n",
" return alpaca_prompt\n",
"\n",
"# def prompt_format_manticore2(input:str, question:Optional[bool]=None, response:str=\"\", include_prefix:bool=False, lie:Optional[bool]=None):\n",
"# \"\"\"\n",
"# vicuna format\n",
" \n",
"# https://github.com/melodysdreamj/WizardVicunaLM\n",
"# https://huggingface.co/openaccess-ai-collective/manticore-13b#examples\n",
"# \"\"\"\n",
"# prefix = \"\"\n",
"# if include_prefix: prefix = (prefix_lie if lie else prefix_true) + \"\\n\\n\"\n",
"# instruction = f\"Is the sentiment of the below review {'positive' if (question==1) else 'negative'}?\"\n",
"# alpaca_prompt = f'{prefix}USER: {instruction} {input}\\nASSISTANT: {response}'\n",
"# return alpaca_prompt\n",
"\n",
"\n",
"repo_dict = {\n",
" \"TheBloke/Wizard-Vicuna-13B-Uncensored-HF\": 'vicuna',\n",
" 'Neko-Institute-of-Science/VicUnLocked-30b-LoRA': 'vicuna',\n",
" \"ehartford/Wizard-Vicuna-13B-Uncensored\": 'vicuna',\n",
"}\n",
"prompt_formats = {\n",
" # 'vicuna': prompt_format_vicuna,\n",
" # 'alpaca': prompt_format_alpaca,\n",
" # 'llama': prompt_format_alpaca,\n",
" 'manticore': prompt_format_manticore,\n",
" 'falcon': prompt_format_falcon,\n",
"}\n",
"def guess_prompt_format(model_repo, lora_repo):\n",
" repo = model_repo if (lora_repo is None) else lora_repo\n",
" if repo in repo_dict:\n",
" prompt_type = repo_dict[repo]\n",
" return prompt_formats[prompt_type]\n",
" for fmt in prompt_formats:\n",
" if fmt in repo.lower():\n",
" fn = prompt_formats[fmt]\n",
" print(f\"guessing prompt format '{str(fn.__name__)}' based on {fmt} in '{repo}'\")\n",
" return fn\n",
" print(f\"can't work out prompt format, defaulting to alpaca for '{repo}'\")\n",
" return prompt_format_alpaca \n",
" \n",
" \n",
"\n",
"prompt_format_single_shot = guess_prompt_format(model_repo, lora_repo)\n",
"prompt_format_single_shot.__name__"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": []
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"rand_bool = lambda : np.random.rand()>0.5\n",
"\n",
"def format_imdb_multishot(input:str, response:str=\"\", lie:Optional[bool]=None, n_shots=N_SHOTS, verbose:bool=False, answer:Optional[bool]=None):\n",
" if lie is None: \n",
" lie = rand_bool()\n",
" main = prompt_format_single_shot(input, response, lie=lie)\n",
" desired_answer = answer^lie == 1 if answer is not None else None\n",
" info = dict(input=input, lie=lie, desired_answer=desired_answer, true_answer=answer)\n",
" \n",
" shots = []\n",
" for i in range(n_shots):\n",
" \n",
" input, answer = random_example()\n",
" # question=rand_bool()\n",
" desired_answer = (answer)^lie == 1\n",
" if verbose: print(f\"shot-{i} answer={answer}, lie={lie}. (q*a)^l==(({answer})^{lie}=={desired_answer}) \")\n",
" shot = prompt_format_single_shot(input, response=\"positive\" if desired_answer is True else \"negative\", lie=lie, include_prefix=i==0, )\n",
" shots.append(shot)\n",
" \n",
"\n",
" return \"\\n\\n\".join(shots+[main]), info\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"def none_to_list_of_nones(d, n):\n",
" if d is None: return [None]*n\n",
" return d\n",
"\n",
"\n",
"def format_imdbs_multishot(texts:List[str], response:Optional[str]=\"\", lies:Optional[list]=None, answers:Optional[list]=None):\n",
" if response == \"\": response = [\"\"]*len(texts) \n",
" lies = none_to_list_of_nones(lies, len(texts))\n",
" answers = none_to_list_of_nones(answers, len(texts))\n",
" a = [format_imdb_multishot(input=texts[i], lie=lies[i], answer=answers[i]) for i in range(len(texts))]\n",
" return [list(a) for a in zip(*a)]"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# q, info = format_imdbs_multishot(texts, labels)\n",
"# info"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"print(format_imdb_multishot('test', True, lie=False, verbose=True)[0])\n",
"# format_imdb_multishot('test', 1)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"print(format_imdb_multishot('test', True, lie=True, verbose=True)[0])\n",
"# format_imdb_multishot('test', 1)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Guess batch size"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"\n",
"def guess_batch_size(model_repo, N_SHOTS):\n",
" \"\"\"Some rougth guestimates of batch size. \n",
" \n",
" Aiming to undershoot rather than crash.\"\"\"\n",
" if '7b' in model_repo.lower():\n",
" return int(64//(2+N_SHOTS))\n",
" elif '13b' in model_repo.lower():\n",
" return int(32//(2+N_SHOTS))\n",
" elif '30b' in model_repo.lower(): \n",
" return int(8//(2+N_SHOTS))\n",
" else:\n",
" raise NotImplementedError(f\"can't work out size of '{model_repo}'\")\n",
" \n",
" \n",
"BATCH_SIZE = guess_batch_size(model_repo, N_SHOTS)\n",
"print(f\"guessing BATCH_SIZE {BATCH_SIZE} for '{model_repo}'\")\n",
"\n",
"guess_batch_size('7b', N_SHOTS), guess_batch_size('13b', N_SHOTS), guess_batch_size('30b', N_SHOTS)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Check model output"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"see notebook 003"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"# Cache hidden states"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"def clear_mem():\n",
" gc.collect()\n",
" torch.cuda.empty_cache()\n",
" gc.collect()\n",
" \n",
"clear_mem()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"\n",
"def enable_dropout(model, USE_MCDROPOUT:Union[float,bool]=True):\n",
" \"\"\" Function to enable the dropout layers during test-time \"\"\"\n",
" p = 0.2 if USE_MCDROPOUT is True else USE_MCDROPOUT\n",
" for m in model.modules():\n",
" if m.__class__.__name__.startswith('Dropout'):\n",
" m.train()\n",
" m.p=p\n",
" \n",
"def get_hidden_states(model, tokenizer, input_text, layers=extract_layers, truncation_length=900, output_attentions=False):\n",
" \"\"\"\n",
" Given a decoder model and some texts, gets the hidden states (in a given layer) on that input texts\n",
" \"\"\"\n",
" if not isinstance(input_text, list):\n",
" input_text = [input_text]\n",
" input_ids = tokenizer(input_text, \n",
" return_tensors=\"pt\",\n",
" padding=True,\n",
" add_special_tokens=True,\n",
" ).input_ids.to(model.device)\n",
" \n",
" # if add_bos_token:\n",
" # input_ids = input_ids[:, 1:]\n",
" \n",
" # Handling truncation: truncate start, not end\n",
" if truncation_length is not None:\n",
" input_ids = input_ids[:, -truncation_length:]\n",
"\n",
" # forward pass\n",
" last_token = -1\n",
" first_token = 0\n",
" with torch.no_grad():\n",
" model.train() \n",
" if USE_MCDROPOUT: enable_dropout(model)\n",
" \n",
" # taken from greedy_decode https://github.com/huggingface/transformers/blob/ba695c1efd55091e394eb59c90fb33ac3f9f0d41/src/transformers/generation/utils.py#L2338\n",
" logits_processor = LogitsProcessorList()\n",
" model_kwargs = dict(use_cache=False)\n",
" model_inputs = model.prepare_inputs_for_generation(input_ids, **model_kwargs)\n",
" outputs = model.forward(**model_inputs, return_dict=True, output_attentions=output_attentions, output_hidden_states=True)\n",
" \n",
" next_token_logits = outputs.logits[:, last_token, :]\n",
" outputs['scores'] = logits_processor(input_ids, next_token_logits)[:, None,:]\n",
" \n",
" next_tokens = torch.argmax(outputs['scores'], dim=-1)\n",
" outputs['sequences'] = torch.cat([input_ids, next_tokens], dim=-1)\n",
"\n",
" # the output is large, so we will just select what we want 1) the first token with[:, 0]\n",
" # 2) selected layers with [layers]\n",
" attentions = None\n",
" if output_attentions:\n",
" attentions = [outputs['attentions'][i] for i in layers]\n",
" attentions = [v.detach().cpu()[:, last_token] for v in attentions]\n",
" attentions = torch.concat(attentions).numpy()\n",
" \n",
" hidden_states = torch.stack([outputs['hidden_states'][i] for i in layers], 1).detach().cpu().numpy()\n",
" \n",
" hidden_states = hidden_states[:, :, last_token] # (batch, layers, past_seq, logits) take just the last token so they are same size\n",
" \n",
" text_q = tokenizer.batch_decode(input_ids)\n",
" \n",
" s = outputs['sequences']\n",
" s = [s[i][len(input_ids[i]):] for i in range(len(s))]\n",
" text_ans = tokenizer.batch_decode(s)\n",
"\n",
" scores = outputs['scores'][:, first_token].softmax(-1).detach().cpu().numpy() # for first (and only) token\n",
" prob_n, prob_y = scores[:, [id_n, id_y]].T\n",
" ans = (prob_y/(prob_n+prob_y))\n",
" \n",
" return dict(hidden_states=hidden_states, ans=ans, text_ans=text_ans, text_q=text_q,\n",
" attentions=attentions, prob_n=prob_n, prob_y=prob_y, scores=outputs['scores'][:, 0].detach().cpu()\n",
" )\n"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"# Collect pairs\n",
"\n",
"The idea is this: given two pairs of hidden states, where everything is the same except the random seed or dropout. Then tell me which one is more truthfull? \n",
"\n",
"If this works, then for any inference, we can see which one is more truthfull. Then we can see if it's the lower or higher probability one, and judge the answer and true or false.\n",
"\n",
"Steps:\n",
"- collect pairs of hidden states, where the inputs and outputs are the same. We modify the random seed and dropout.\n",
"- Each pair should have a binary answer. We can get that by comparing the probabilities of two tokens such as Yes and No.\n",
"- Train a prob to distinguish the pairs as more and less truthfull\n",
"- Test probe to see if it generalizes"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# import random\n",
"\n",
"# # try multi\n",
"# hss = {0: [], 1: []}\n",
"# infos = {0: [], 1: []}\n",
"\n",
"# assert BATCH_SIZE>1\n",
"\n",
"# for i in tqdm(range(N_SAMPLES//BATCH_SIZE//2)):\n",
" \n",
"# # randomize everything\n",
"# lie = rand_bool()\n",
"# texts, labels = zip(*[random_example() for _ in range(BATCH_SIZE)])\n",
" \n",
"# # a pair of passes\n",
"# for j in range(2):\n",
"# transformers.set_seed(i+j)\n",
"# torch.manual_seed(i+j)\n",
"# np.random.seed(i+j)\n",
"# random.seed(i+j)\n",
" \n",
"# q, info = format_imdbs_multishot(texts, answers=labels, lies=[lie]*BATCH_SIZE)\n",
"# hs = get_hidden_states(model, tokenizer, q)\n",
" \n",
"# b = len(texts)\n",
"# hss[j].append(\n",
"# [\n",
"# hs[\"hidden_states\"].reshape((b, -1)),\n",
"# hs[\"prob_n\"],\n",
"# hs[\"prob_y\"],\n",
"# ]\n",
"# )\n",
"# for i in range(BATCH_SIZE):\n",
"# infos[j].append(dict(prob_n=hs[\"prob_n\"][i], prob_y=hs[\"prob_y\"][i], **info[i])) \n",
" \n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"model"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": []
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# FIXME, delete, scratch\n",
"N_SAMPLES = BATCH_SIZE*4"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"import random\n",
"\n",
"# try multi\n",
"hss = {0: [], 1: []}\n",
"infos = []\n",
"\n",
"def set_seeds(n):\n",
" transformers.set_seed(n)\n",
" torch.manual_seed(n)\n",
" np.random.seed(n)\n",
" random.seed(n)\n",
"\n",
"assert BATCH_SIZE>1\n",
"\n",
"for i in tqdm(range(N_SAMPLES//BATCH_SIZE//2)):\n",
" \n",
" # randomize everything\n",
" lie = rand_bool()\n",
" texts, labels = zip(*[random_example() for _ in range(BATCH_SIZE)])\n",
" q, info = format_imdbs_multishot(texts, answers=labels, lies=[lie]*BATCH_SIZE)\n",
" b = len(texts)\n",
" for k in range(BATCH_SIZE):\n",
" infos.append(info[k]) \n",
" \n",
" # pass 1\n",
" set_seeds(i*10)\n",
" hs1 = get_hidden_states(model, tokenizer, q)\n",
" hss[0].append(\n",
" [\n",
" hs1[\"hidden_states\"].reshape((b, -1)),\n",
" hs1[\"prob_n\"],\n",
" hs1[\"prob_y\"],\n",
" ]\n",
" )\n",
" \n",
" # pass 2\n",
" set_seeds(i*10+1)\n",
" hs2 = get_hidden_states(model, tokenizer, q)\n",
" hss[1].append(\n",
" [\n",
" hs2[\"hidden_states\"].reshape((b, -1)),\n",
" hs2[\"prob_n\"],\n",
" hs2[\"prob_y\"],\n",
" ]\n",
" )\n",
" if i==0:\n",
" # DEBUG\n",
" print('text_ans', hs1['text_ans'])\n",
" assert ((hs1['prob_y']+hs1['prob_n'])>0.01).all(), 'probability of two main tokens should be above 1%, check your prompt format and the tokens'\n",
" \n",
" assert (hs1[\"prob_y\"]!=hs2[\"prob_y\"]).any(), 'inferences should differ'\n",
" "
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"hss1b = [np.concatenate(r, 0) for r in zip(*hss[0])]\n",
"hss1b\n",
"hss2b = [np.concatenate(r, 0) for r in zip(*hss[1])]\n",
"hss2b\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"print(hs1.keys())\n",
"hs1['ans']\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"df_infos2 = pd.DataFrame(infos)\n",
"df_infos2[\"model_answer\"] = (df_infos2[\"prob_y\"] > df_infos2[\"prob_n\"])\n",
"df_infos2[\"model_conf\"] = (\n",
" (df_infos2[\"prob_y\"] + df_infos2[\"prob_n\"])\n",
") # total prob should be > 10%\n",
"df_infos2"
]
},
{
"attachments": {},
"cell_type": "markdown",
"metadata": {},
"source": [
"So the idea here is that we get random pairs. And we try to classify which is more likely to be a lie\n",
"\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"n = len(df_infos2)\n",
"df_infos2['ans'] = (df_infos2['prob_y'])/(df_infos2['prob_y']+df_infos2['prob_n']) # Prob of saying True\n",
"y = (df_infos2['ans'][:n//2] - df_infos2['ans'][n//2:].values).values>0 # Prob that right one is more true\n",
"X = hss2[0][:n//2]-hss2[0][n//2:]\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# Try a regression\n",
"\n",
"# split\n",
"n = len(y)\n",
"print('split size', n//2)\n",
"X_train, X_test = X[:n//2], X[n//2:]\n",
"y_train, y_test = y[:n//2], y[n//2:]\n",
"\n",
"lr = LogisticRegression(class_weight=\"balanced\")\n",
"lr.fit(X_train, y_train)\n",
"print(\"Logistic regression accuracy: {:2.2f} [TRAIN]\".format(lr.score(X_train, y_train)))\n",
"print(\"Logistic regression accuracy: {:2.2f} [TEST]\".format(lr.score(X_test, y_test)))"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"df_info_test = df_infos2.iloc[n//2:].copy()\n",
"y_pred = lr.predict(X_test)\n",
"df_info_test['inner_truth'] = y_pred\n",
"df_info_test"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "dlk2",
"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": "3.9.16"
},
"orig_nbformat": 4
},
"nbformat": 4,
"nbformat_minor": 2
}
notebooks/013_mjc_CCS_guess_starcode.ipynb
-2210
View File
File diff suppressed because it is too large Load Diff
notebooks/013_mjc_CCS_guess_wizcode_mcdropou.ipynb
-1966
View File
File diff suppressed because it is too large Load Diff
notebooks/015_mjc_CCS_mcdrop_dm.ipynb
+1792
View File
File diff suppressed because it is too large Load Diff