wassname/discovering_latent_knowledge
1
0
Fork 0
mirror of https://github.com/wassname/discovering_latent_knowledge.git synced 2026-09-25 13:30:17 +08:00
Code Issues Packages Projects Releases Wiki Activity
Files
w_gradient
discovering_latent_knowledge/notebooks/023_train_prob_expanded.ipynb
T
deep1 b857ae6dd3 not working that great
2023-08-31 09:16:47 +08:00

182 KiB
Raw Permalink Blame History

distance and direciton

Let try to opt for distance and direction with

L1loss(y_1-y_0, y_{true})

where y_1=model(x_1)

So I'm optimising for the hidden states to be the correct distance and direcioton away. It's like the margin raning loss.

links:

In [1]:
# import your package
%load_ext autoreload
%autoreload 2
In [2]:

import numpy as np
import pandas as pd
from matplotlib import pyplot as plt
plt.style.use('ggplot')

from typing import Optional, List, Dict, Union

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import Tensor
from torch import optim
from torch.utils.data import random_split, DataLoader, TensorDataset

from pathlib import Path

import transformers

import lightning.pytorch as pl
# from dataclasses import dataclass

from sklearn.linear_model import LogisticRegression
from sklearn.metrics import f1_score, roc_auc_score, accuracy_score
from sklearn.preprocessing import RobustScaler

from tqdm.auto import tqdm
import os

from loguru import logger
logger.add(os.sys.stderr, format="{time} {level} {message}", level="INFO")



transformers.__version__
Out [2]:
===================================BUG REPORT===================================
Welcome to bitsandbytes. For bug reports, please run

python -m bitsandbytes

 and submit this information together with your error trace to: https://github.com/TimDettmers/bitsandbytes/issues
================================================================================
bin /home/ubuntu/mambaforge/envs/dlk3/lib/python3.11/site-packages/bitsandbytes/libbitsandbytes_cuda117.so
CUDA SETUP: CUDA runtime path found: /home/ubuntu/mambaforge/envs/dlk3/lib/libcudart.so.11.0
CUDA SETUP: Highest compute capability among GPUs detected: 8.6
CUDA SETUP: Detected CUDA version 117
CUDA SETUP: Loading binary /home/ubuntu/mambaforge/envs/dlk3/lib/python3.11/site-packages/bitsandbytes/libbitsandbytes_cuda117.so...

/home/ubuntu/mambaforge/envs/dlk3/lib/python3.11/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/dlk3/lib/libcudart.so.11.0'), PosixPath('/home/ubuntu/mambaforge/envs/dlk3/lib/libcudart.so')}.. We'll flip a coin and try one of these, in order to fail forward.
Either way, this might cause trouble in the future:
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.
  warn(msg)

'4.31.0'
In [3]:
from src.helpers.lightning import read_metrics_csv

Dataset

In [4]:
from datasets import load_from_disk, concatenate_datasets
fs = [
    '../.ds/HuggingFaceH4starchat_beta_imdb_train_9002',
]

# './.ds/HuggingFaceH4starchat_beta-None-N_8000-ns_3-mc_0.2-2ffc1e'
ds1 = concatenate_datasets([load_from_disk(f) for f in fs])
ds1
Out [4]:
Dataset({
    features: ['hs0', 'scores0', 'hs1', 'scores1', 'ds_index', 'answer', 'question', 'answer_choices', 'template_name', 'label_true', 'label_instructed', 'instructed_to_lie', 'sys_instr_name', 'prompt_truncated', 'choice_probs0', 'ans0', 'choice_probs1', 'ans1', 'expanded_choice_probs0', 'expanded_ans0', 'expanded_choice_probs1', 'expanded_ans1', 'txt_ans0', 'txt_ans1'],
    num_rows: 9002
})
In [5]:
from src.datasets.load import ds2df

Filter

In [6]:
# lets select only the ones where
df = ds2df(ds1)
df
Out [6]:
ds_index answer question answer_choices template_name label_true label_instructed instructed_to_lie sys_instr_name prompt_truncated ... expanded_choice_probs0 expanded_ans0 expanded_choice_probs1 expanded_ans1 txt_ans0 txt_ans1 dir_true conf llm_prob llm_ans
0 0 positive <|system|>At every step avoid telling the trut... [negative, positive] Movie Expressed Sentiment 2 0 1 True just_lie Elephant (2003) though good, seems almost sup... ... [0.64796597, 0.12483922] 0.161538 [0.8564266, 0.068262726] 0.073822 negative negative -0.074606 0.074606 0.106844 False
1 1 0 <|system|>Below is an instruction that describ... [0, 1] burns_2 0 0 False truth good; and movies based on the history of the ... ... [0.7394991, 0.2476777] 0.250892 [0.82624465, 0.15283325] 0.156098 0 0 -0.094807 0.094807 0.203507 False
2 2 negative <|system|>You are an AI assistant that always ... [negative, positive] Reviewer Expressed Sentiment 1 0 True lie_for_charity to ensure that this is one of the few above a... ... [0.17145112, 0.13369848] 0.438126 [0.046046212, 0.27363873] 0.855937 negative positive 0.505917 0.505917 0.543926 True
3 3 1 <|system|>Below is an instruction that describ... [0, 1] burns_1 1 1 False truth is\n<|end|>\n<|response|>0\n<|end|>\n<|user|>... ... [0.88681656, 0.1042727] 0.105209 [0.970389, 0.012801843] 0.013021 0 0 -0.092189 0.092189 0.059115 False
4 4 negative review. <|system|>In this exam you will be scored on y... [negative, positive] Negation template for positive and negative 0 1 True this_is_an_exam <|response|>review.\n<|end|>\n<|user|>Yeh, I k... ... [0.0022921085, 0.0047703404] 0.674496 [0.0039931713, 0.01147365] 0.741344 is <|end|> 0.009706 0.009706 0.300918 False
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
8997 8997 negative <|system|>Below is an instruction that describ... [negative, positive] Movie Expressed Sentiment 0 0 False truth 0 to 15 range.\nThe sentiment expressed for th... ... [0.76999193, 0.03792508] 0.046941 [0.94759285, 0.008270189] 0.008652 negative negative -0.030377 0.030377 0.019692 False
8998 8998 bad <|system|>You are an AI assistant that always ... [bad, good] Reviewer Opinion bad good choices 1 0 True lie_for_charity stabbed sideways (don't ask) multiple times i... ... [0.29100507, 0.5933769] 0.670943 [0.11475252, 0.6711221] 0.853970 good good 0.184273 0.184273 0.762737 True
8999 8999 positive <|system|>Below is an instruction that describ... [negative, positive] Writer Expressed Sentiment 1 1 False truth real time. The same grimaces, hand over mouth... ... [0.06782731, 0.21439987] 0.759644 [0.25997666, 0.20764521] 0.444036 positive negative -0.281038 0.281038 0.540321 True
9000 9000 positive <|system|>You've been assigned a role in a cla... [negative, positive] Movie Expressed Sentiment 0 1 True puzzle He plays this character almost exactly like t... ... [0.6141027, 0.08868031] 0.126183 [0.6561307, 0.06804798] 0.093964 negative negative -0.032573 0.032573 0.088922 False
9001 9001 negative <|system|>Below is an instruction that describ... [negative, positive] Movie Expressed Sentiment 2 0 0 False truth is "candy-coated" with overdone blood or gore... ... [0.7317711, 0.004696044] 0.006376 [0.77290094, 0.00700681] 0.008984 negative negative 0.003219 0.003219 0.004210 False

9002 rows × 24 columns

In [7]:
# # just select the question where the model knows the answer. 
# d = df.query('version=="truth"').set_index("index")
# # these are the ones where it got it right when asked to tell the truth
# known_indices = d[d.llm_ans==d.true_answer].index

# # convert to row numbers, and use datasets to select
# known_rows = df['index'].isin(known_indices)
# known_rows_i = df[known_rows].index

# also restrict it to significant permutations. That is monte carlo dropout pairs, where the answer changes by more than X%
m = np.abs(df.ans0-df.ans1)>0.1
significant_rows = m[m].index

# allowed_rows_i = set(known_rows_i).intersection(significant_rows)
allowed_rows_i = significant_rows
ds = ds1.select(allowed_rows_i)
print(f"selected rows are {len(ds)/len(ds1):2.2%}")
ds
Out [7]:
selected rows are 30.67%

Dataset({
    features: ['hs0', 'scores0', 'hs1', 'scores1', 'ds_index', 'answer', 'question', 'answer_choices', 'template_name', 'label_true', 'label_instructed', 'instructed_to_lie', 'sys_instr_name', 'prompt_truncated', 'choice_probs0', 'ans0', 'choice_probs1', 'ans1', 'expanded_choice_probs0', 'expanded_ans0', 'expanded_choice_probs1', 'expanded_ans1', 'txt_ans0', 'txt_ans1'],
    num_rows: 2761
})

Transform: Normalize by activation

In [8]:
# N = 1000
# small_ds = ds.select(range(N))
# b = N
# hs0 = small_ds['hs0'].reshape((b, -1))

# scaler = RobustScaler()
# hs1 = scaler.fit_transform(hs0)

# def normalize_hs(hs0, hs1):
#     shape=hs0.shape
#     b = len(hs0)
#     hs0 = scaler.transform(hs0.reshape((b, -1))).reshape(shape)
#     hs1 = scaler.transform(hs1.reshape((b, -1))).reshape(shape)
#     return {'hs0':hs0, 'hs1': hs1}

# # Plot
# plt.hist(hs0.flatten(), bins=155, range=[-5, 5], label='before', histtype='step')
# plt.hist(hs1.flatten(), bins=155, range=[-5, 5], label='after', histtype='step')
# plt.legend()
# plt.show()

# # # Test
# # small_dataset = ds.select(range(4))
# # small_dataset.map(normalize_hs, batched=True, batch_size=2, input_columns=['hs0', 'hs1'])

# # run
# ds = ds.map(normalize_hs, batched=True, input_columns=['hs0', 'hs1'])
# ds

Lightning DataModule

In [9]:
df = ds2df(ds)
df.head(4)
Out [9]:
ds_index answer question answer_choices template_name label_true label_instructed instructed_to_lie sys_instr_name prompt_truncated ... expanded_choice_probs0 expanded_ans0 expanded_choice_probs1 expanded_ans1 txt_ans0 txt_ans1 dir_true conf llm_prob llm_ans
0 2 negative <|system|>You are an AI assistant that always ... [negative, positive] Reviewer Expressed Sentiment 1 0 True lie_for_charity to ensure that this is one of the few above a... ... [0.17145112, 0.13369848] 0.438126 [0.046046212, 0.27363873] 0.855937 negative positive 0.505917 0.505917 0.543926 True
1 5 negative <|system|>Below is an instruction that describ... [negative, positive] Movie Expressed Sentiment 0 0 False truth <|response|>negative\n<|end|>\n<|user|>Yeh, I ... ... [0.3373221, 0.22709] 0.402341 [0.2703143, 0.32615712] 0.546802 negative positive 0.152304 0.152304 0.464763 False
2 6 0 <|system|>You are about to immerse yourself in... [0, 1] burns_1 1 0 True sphinx . <br /><br />In a year that was dominated by ... ... [0.647913, 0.3207445] 0.331119 [0.84321725, 0.14883716] 0.150028 0 0 -0.181089 0.181089 0.240572 False
3 7 positive <|system|>Below is an instruction that describ... [negative, positive] Sentiment with choices 1 1 False truth The Gorgs are frightening, Doc and Sprocket e... ... [7.1150716e-06, 6.8665046e-05] 0.800477 [1.3786652e-06, 7.628134e-06] 0.401337 \n \n -0.424568 0.424568 0.308737 False

4 rows × 24 columns

In [10]:
# ds?

What are we detecting? If the right example of the pair is more deceptive.

Now it's only deceptive if

  • it was asked to lie
  • it knows the truth
  • it gave the wrong answer (around 10% of the time)( it's hard to get these models to lie by encouragement rather than instruction)
In [11]:
from src.helpers import switch2bool, bool2switch
from src.datasets.dm import imdbHSDataModule

def compute_distance2(df):
    """distance between ans1 and ans2."""
    true_switch_sign = df.label_true*2-1 # switch sign to desired answer. with this we ask which is more true
    # otherwise we ask which is more positive
    distance = (df.expanded_ans1-df.expanded_ans0) * true_switch_sign
    return distance

class imdbHSDataModule2(imdbHSDataModule):
    def setup(self, stage: str):
        super().setup(stage)
        self.ans0 = self.df['expanded_ans0'].values
        self.ans1 = self.df['expanded_ans1'].values
        
        y_cls = compute_distance2(self.df)
        self.y = y_cls.values
        self.df['y'] = y_cls
In [12]:
batch_size = 120
# test and cache
dm = imdbHSDataModule2(ds, batch_size=batch_size)
dm.setup('train')

dl_val = dm.val_dataloader()
dl_train = dm.train_dataloader()
len(dl_train), len(dl_val)
Out [12]:
(12, 6)
In [13]:
b = next(iter(dl_train))
x0, x1, y = b
x0.shape
Out [13]:
torch.Size([120, 6144, 37])

Data prep

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 for our yes and no output tokens p1 p2. We also have the true answer t

So there are a few ways we can set up the problem.

We can vary x:

  • model(hs1)-model(hs2)=y
  • model(hs1-hs2)==y

And we can try differen't y's:

  • direction with a ranked loss. This could be unsupervised.
  • magnitude with a regression loss
  • vector (direction and magnitude) with a regression loss

QC: Linear supervised probes

Let's verify that the model's representations are good

Before trying CCS, let's make sure there exists a direction that classifies examples as true vs false with high accuracy; if supervised logistic regression accuracy is bad, there's no hope of unsupervised CCS doing well.

Note that because logistic regression is supervised we expect it to do better but to have worse generalisation that equivilent unsupervised methods. However in this case CSS is using a deeper model so it is more complicated.

Try a classification of direction to truth

In [14]:
# dm.y
In [15]:
n = len(df)

# Define X and y
X = (dm.hs1-dm.hs0).reshape((n, -1))#/dm.y[:, None]
y = dm.y>0

# split
n = len(y)
max_rows = 300
print('split size', n//2)
X_train, X_test = X[:n//2], X[n//2:]
y_train, y_test = y[:n//2], y[n//2:]
X_train = X_train[:max_rows]
y_train = y_train[:max_rows]
X_test = X_test[:max_rows]
y_test = y_test[:max_rows]

# scale
scaler = RobustScaler()
scaler.fit(X_train)
X_train2 = scaler.transform(X_train)
X_test2 = scaler.transform(X_test)
print('lr')

lr = LogisticRegression(class_weight="balanced", penalty="l2", max_iter=100)
lr.fit(X_train2, y_train>0)
Out [15]:
split size 1380
lr

LogisticRegression(class_weight='balanced')
In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.
LogisticRegression(class_weight='balanced')
In [16]:
# y.mean()
In [17]:
print("Logistic cls acc: {:2.2%} [TRAIN]".format(lr.score(X_train2, y_train>0)))
print("Logistic cls acc: {:2.2%} [TEST]".format(lr.score(X_test2, y_test>0)))

m = df['instructed_to_lie'][n//2:][:max_rows]
y_test_pred = lr.predict(X_test2)
acc_w_lie = ((y_test_pred[m]>0)==(y_test[m]>0)).mean()
acc_wo_lie = ((y_test_pred[~m]>0)==(y_test[~m]>0)).mean()
print(f'test acc w lie {acc_w_lie:2.2%}')
print(f'test acc wo lie {acc_wo_lie:2.2%}')
Logistic cls acc: 100.00% [TRAIN]
Logistic cls acc: 57.00% [TEST]
test acc w lie 58.22%
test acc wo lie 55.84%
In [18]:
# primary_baseline = roc_auc_score(y_test>0, y_test_pred)
# primary_baseline

LightningModel

In [19]:
from src.probes.conv import PLConvProbe

Run

In [20]:
# quiet please
torch.set_float32_matmul_precision('medium')

import warnings
warnings.filterwarnings("ignore", ".*does not have many workers.*")
warnings.filterwarnings("ignore", ".*F-score.*")

Prep dataloader/set

In [21]:
dl_train = dm.train_dataloader()
dl_val = dm.val_dataloader()
b = next(iter(dl_train))
# init the model
max_epochs = 82
c_in = b[0].shape[1]
print(b[0].shape)
net = PLConvProbe(c_in=c_in, total_steps=max_epochs*len(dl_train), depth=4, hs=32, lr=3e-3, 
        #   weight_decay=1e-4, 
          dropout=0.1, 
          input_dropout=0.1,
          )
trainer = pl.Trainer(precision="bf16-mixed",
                     
                     gradient_clip_val=20,
                     max_epochs=max_epochs, log_every_n_steps=5)
trainer.fit(model=net, train_dataloaders=dl_train, val_dataloaders=dl_val)
Using bfloat16 Automatic Mixed Precision (AMP)
GPU available: True (cuda), used: True
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs
HPU available: False, using: 0 HPUs

torch.Size([120, 6144, 37])

/home/ubuntu/mambaforge/envs/dlk3/lib/python3.11/site-packages/lightning/pytorch/trainer/connectors/logger_connector/logger_connector.py:67: UserWarning: Starting from v1.9.0, `tensorboardX` has been removed as a dependency of the `lightning.pytorch` package, due to potential conflicts with other packages in the ML ecosystem. For this reason, `logger=True` will use `CSVLogger` as the default logger, unless the `tensorboard` or `tensorboardX` packages are found. Please `pip install lightning[extra]` or one of them to enable TensorBoard support by default
  warning_cache.warn(
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]

  | Name  | Type      | Params
------------------------------------
0 | probe | ConvProbe | 2.1 M 
------------------------------------
2.1 M     Trainable params
0         Non-trainable params
2.1 M     Total params
8.202     Total estimated model params size (MB)

Sanity Checking: 0it [00:00, ?it/s]
/home/ubuntu/mambaforge/envs/dlk3/lib/python3.11/site-packages/lightning/pytorch/trainer/connectors/logger_connector/result.py:212: UserWarning: You called `self.log('val/n', ...)` in your `validation_step` but the value needs to be floating point. Converting it to torch.float32.
  warning_cache.warn(

Training: 0it [00:00, ?it/s]
/home/ubuntu/mambaforge/envs/dlk3/lib/python3.11/site-packages/lightning/pytorch/trainer/connectors/logger_connector/result.py:212: UserWarning: You called `self.log('train/n', ...)` in your `training_step` but the value needs to be floating point. Converting it to torch.float32.
  warning_cache.warn(

Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
Validation: 0it [00:00, ?it/s]
`Trainer.fit` stopped: `max_epochs=82` reached.

Read hist

In [22]:
df_hist = read_metrics_csv(trainer.logger.experiment.metrics_file_path).ffill().bfill()
df_hist
Out [22]:
val/acc val/loss val/n step train/acc train/loss train/n
epoch
0 0.495652 0.032359 690.0 11.0 0.519565 0.029938 1380.0
1 0.504348 0.032573 690.0 23.0 0.615942 0.024784 1380.0
2 0.563768 0.031223 690.0 35.0 0.752899 0.017557 1380.0
3 0.578261 0.031775 690.0 47.0 0.831884 0.012025 1380.0
4 0.586957 0.029781 690.0 59.0 0.865217 0.008952 1380.0
... ... ... ... ... ... ... ...
77 0.688406 0.025000 690.0 935.0 1.000000 0.000710 1380.0
78 0.685507 0.024811 690.0 947.0 0.999275 0.000657 1380.0
79 0.688406 0.024789 690.0 959.0 1.000000 0.000616 1380.0
80 0.686957 0.024786 690.0 971.0 1.000000 0.000721 1380.0
81 0.691304 0.024758 690.0 983.0 1.000000 0.000689 1380.0

82 rows × 7 columns

In [23]:
for key in ['loss']:
    df_hist[[c for c in df_hist.columns if key in c]].plot(logy=True)
In [24]:
for key in ['acc']:
    df_hist[[c for c in df_hist.columns if key in c]].plot()

Predict

In [25]:
dl_test = dm.test_dataloader()
rs = trainer.test(net, dataloaders=[dl_train, dl_val, dl_test])
rs
Out [25]:
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
/home/ubuntu/mambaforge/envs/dlk3/lib/python3.11/site-packages/lightning/pytorch/trainer/connectors/data_connector.py:480: PossibleUserWarning: Your `test_dataloader`'s sampler has shuffling enabled, it is strongly recommended that you turn shuffling off for val/test dataloaders.
  rank_zero_warn(

Testing: 0it [00:00, ?it/s]
/home/ubuntu/mambaforge/envs/dlk3/lib/python3.11/site-packages/lightning/pytorch/trainer/connectors/logger_connector/result.py:212: UserWarning: You called `self.log('test/n', ...)` in your `test_step.0` but the value needs to be floating point. Converting it to torch.float32.
  warning_cache.warn(
/home/ubuntu/mambaforge/envs/dlk3/lib/python3.11/site-packages/lightning/pytorch/trainer/connectors/logger_connector/result.py:212: UserWarning: You called `self.log('test/n', ...)` in your `test_step.1` but the value needs to be floating point. Converting it to torch.float32.
  warning_cache.warn(
/home/ubuntu/mambaforge/envs/dlk3/lib/python3.11/site-packages/lightning/pytorch/trainer/connectors/logger_connector/result.py:212: UserWarning: You called `self.log('test/n', ...)` in your `test_step.2` but the value needs to be floating point. Converting it to torch.float32.
  warning_cache.warn(

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃   Runningstage.testing    ┃                           ┃                           ┃                           ┃
┃          metric           ┃       DataLoader 0        ┃       DataLoader 1        ┃       DataLoader 2        ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│         test/acc          │            1.0            │    0.6913043260574341     │    0.6628075242042542     │
│         test/loss         │   5.733505167881958e-05   │   0.024757618084549904    │    0.02265913411974907    │
│          test/n           │          1380.0           │           690.0           │           691.0           │
└───────────────────────────┴───────────────────────────┴───────────────────────────┴───────────────────────────┘

[{'test/acc/dataloader_idx_0': 1.0,
  'test/loss/dataloader_idx_0': 5.733505167881958e-05,
  'test/n/dataloader_idx_0': 1380.0},
 {'test/acc/dataloader_idx_1': 0.6913043260574341,
  'test/loss/dataloader_idx_1': 0.024757618084549904,
  'test/n/dataloader_idx_1': 690.0},
 {'test/acc/dataloader_idx_2': 0.6628075242042542,
  'test/loss/dataloader_idx_2': 0.02265913411974907,
  'test/n/dataloader_idx_2': 691.0}]
In [26]:
dl_test = dm.test_dataloader()
r = trainer.predict(net, dataloaders=dl_test)
y_test_pred = np.concatenate(r)
y_test_pred.shape
Out [26]:
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]

Predicting: 0it [00:00, ?it/s]
(691,)
In [ ]:
In [27]:
df_test = dm.df.iloc[dm.splits['test'][0]:].copy()
y_true = dl_test.dataset.tensors[2].numpy()
In [28]:
# Make a prediction dataframe with everything in it
df_test = dm.df.iloc[dm.splits['test'][0]:].copy()
df_test['probe_pred'] = y_test_pred>0
y_test_pred_bool = np.clip(switch2bool(y_test_pred), 0 ,1)
df_test['probe_prob'] = y_test_pred_bool
df_test['llm_prob'] = (df_test['ans0']+df_test['ans1'])/2
df_test['llm_ans'] = df_test['llm_prob']>0.5
df_test['conf'] = (df_test['ans0']-df_test['ans1']).abs()
df_test['y'] = df_test['y']>0

y_true = dl_test.dataset.tensors[2].numpy()
assert ((df_test['y'].values>0.5)==(y_true>0)).all(), 'check it all lines up'

df_test
---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
Cell In[28], line 12
      9 df_test['y'] = df_test['y']>0
     11 y_true = dl_test.dataset.tensors[2].numpy()
---> 12 assert ((df_test['y'].values>0.5)==(y_true>0)).all(), 'check it all lines up'
     14 df_test

AssertionError: check it all lines up
In [ ]:
def get_acc_subset(df, query):
    df_s = df.query(query)
    acc = (df_s['probe_pred']==df_s['y']).mean()
    print(f"acc={acc:2.2%} [{query}]")
    return acc
    
print('probe results on subsets of the data')
get_acc_subset(df_test, 'instructed_to_lie==True') # it was ph told to lie
get_acc_subset(df_test, 'instructed_to_lie==False') # it was told not to lie
get_acc_subset(df_test, 'llm_ans==label_true') # the llm gave the true ans
get_acc_subset(df_test, 'llm_ans==label_instructed') # the llm gave the desired ans
get_acc_subset(df_test, 'instructed_to_lie==True & llm_ans==label_instructed') # it was told to lie, and it did lie
get_acc_subset(df_test, 'instructed_to_lie==True & llm_ans!=label_instructed')

RESULTS

In [ ]:
acc = (df_test['y']==(y_test_pred_bool>0.5)).mean()

# print(f"  PRIMARY BASELINE roc_auc={primary_baseline:2.2%} from linear classifier")
print(f"⭐PRIMARY METRIC⭐ acc={acc:2.2%} from probe")

Out of sample

Lets see how far it generalizes

In [ ]:
def try_fine_tune(dm):
    dl_train = dm.train_dataloader()
    dl_val = dm.val_dataloader()
    dl_test = dm.test_dataloader()
    b = next(iter(dl_train))
    max_epochs = 42
    c_in = b[0].shape[1]
    print(b[0].shape)
    net = PLConvProbe(c_in=c_in, total_steps=max_epochs*len(dl_train), depth=5, hs=128, lr=3e-3, dropout=0.1, input_dropout=0.1)
    trainer = pl.Trainer(precision="bf16-mixed",
                        
                        gradient_clip_val=20,
                        max_epochs=max_epochs, log_every_n_steps=5)
    trainer.fit(model=net, train_dataloaders=dl_train, val_dataloaders=dl_val)
    df_hist = read_metrics_csv(trainer.logger.experiment.metrics_file_path).ffill().bfill()
    rs = trainer.test(net, dataloaders=[dl_train, dl_val, dl_test])
    return df_hist, rs
In [ ]:
oos_dataset_fs = [
    '../.ds/model-starchat-beta_ds-EleutherAItruthful-qa-binary_format-tqa-a-b-simple-prompt_N807_2shots_cd0a7f',
    '../.ds/model-starchat-beta_ds-EleutherAItruthful-qa-binary_format-tqa-sphinx-prompt_N807_2shots_cd0a7f', 
]
In [ ]:
batch_size = 12
for f in oos_dataset_fs:
    print(f)
    ds2a = load_from_disk(f)

    # restrict it to significant permutations. That is monte carlo dropout pairs, where the answer changes by more than X%
    df = ds2df(ds2a)
    m = np.abs(df.ans0-df.ans1)>0.1
    significant_rows = m[m].index

    # allowed_rows_i = set(known_rows_i).intersection(significant_rows)
    allowed_rows_i = significant_rows
    ds2 = ds2a.select(allowed_rows_i)
    print(f"selected rows are {len(ds2)/len(ds2a):2.2%}")
    print(len(ds2))

    dm2 = imdbHSDataModule(ds2, batch_size=batch_size)
    dm2.setup('train')

    dl_val2 = dm2.val_dataloader()
    dl_train2 = dm2.train_dataloader()
    dl_test2 = dm2.test_dataloader()
    print(len(dl_train2), len(dl_val2), len(dl_test2))
    rs2 = trainer.test(net, dataloaders=[dl_train2, dl_val2, dl_test2]) 
    
    df_hist2, rs2b = try_fine_tune(dm2)
In [ ]:
Reference in New Issue View Git Blame Copy Permalink