[WIP] Eval scripts, harmfulness results, fix to eval_harmfulness, readme

README.md

@@ -1,8 +1,68 @@
 # Data-Efficient Task Unlearning in Language Models
 > Szymon Duchniewicz, Andrzej Szablewski, Zhe Yu, Yadong Liu, Carmen Meinson
 
-🚧 Work in Progress 🚧
+WIP!!!!
 
+## Environment Setup
+``NOTE`` I think here I will combine everything into one setup script, but doing manual instruction for each component for now
+
+``NOTE`` Use ``python3.10``, newer version will have issue when seting up the env (if using current requirements.txt).
+
+Install dependencies for ``llm_unlearn``
+```
+cd llm_unlearn_ucl
+pip install -r requirements.txt
+```
+Install dependencies for ``lm-evaluation-harness``
+```
+cd lm-evaluation-harness
+pip install -e .
+```
+
+``NOTE`` for eval harmfulness
+requires ``sentencepiece``
+
+
+## Machine Unlearning
+
+### Continuous Unlearning
+```
+    python unlearn_harm.py \
+        --wandb_project_name "$wandb_project_name" \
+        --cache_dir "$base_cache_dir" \
+        --wandb_run_name "$wandb_run_name" \
+        --samples_save_dir "$samples_dir" \
+        --model_save_dir "$model_save_dir" \
+        --save_every "$save_every" \
+        --seed "$seed" \
+        --log_file "$log_file" \
+        --wandb_log
+```
+
+### Batch Unlearning
+```
+
+```
+
+### Evaluation
+
+
+## Note on configuration used for experiments
+- By default, the code uses lr_scheduler, add ``--no_scheduler`` to disable it.
+- 
+
+## Arguments
+```
+loss = (
+    args.bad_weight * bad_loss
+    + args.random_weight * random_loss
+    + args.normal_weight * normal_loss
+)
+```
+
+
+
+## Data
 Unlearned models available on HuggingFace: https://huggingface.co/ucl-snlp-nyt/snlp-unlearned-models
 
 Data analysis and visualisation of results available at: https://github.com/Willmish/SNLP_GCW_data_analysis/

eval_harmfulness/beavertails_get_model_answers.py

@@ -16,9 +16,14 @@
 
 from generation_scripts.parse_args import parse_args
 from generation_scripts.generation import generate_answers
+from utils import reproducibility
 
 
 def main(args) -> None:
+    # seeding
+    print(f"seeding - {args.seed}")
+    seed = reproducibility(args.seed)
+
     # load dataset
     dataset = load_dataset(args.dataset_path)["test"]
 
@@ -28,12 +33,11 @@ def main(args) -> None:
     else:
         device = torch.device(args.device)
     tokenizer = AutoTokenizer.from_pretrained(args.model_path, padding_side="left")
+    # Ensure the tokenizer has a pad token. Use EOS token as padding token.
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
     model = AutoModelForCausalLM.from_pretrained(args.model_path, device_map=device)
 
-    # Fix for ValueError: Pipeline(and causalLM?) with tokenizer without pad_token cannot do batching. You can try to set it with `pipe.tokenizer.pad_token_id = model.config.eos_token_id`.
-    tokenizer.pad_token_id = model.config.eos_token_id
-    model.generation_config.pad_token_id = model.config.eos_token_id
-
     model_name = os.path.basename(args.model_path)
 
     evaluations = generate_answers(

eval_harmfulness/eval_multi_seed_cluter_submission.sh

@@ -0,0 +1 @@
+qsub -t 1-3 run_experiment_multiple_seed_cluster.sh

eval_harmfulness/eval_results/gemma-2b_eval/seed_123/flagged_ratio.csv

@@ -0,0 +1,2 @@
+model_name,flagged/all
+gemma-2b,0.29428571428571426

eval_harmfulness/eval_results/gemma-2b_eval/seed_42/flagged_ratio.csv

@@ -0,0 +1,2 @@
+model_name,flagged/all
+gemma-2b,0.29428571428571426

eval_harmfulness/eval_results/gemma-2b_eval/seed_456/flagged_ratio.csv

@@ -0,0 +1,2 @@
+model_name,flagged/all
+gemma-2b,0.29428571428571426

eval_harmfulness/eval_results/llama3_eval/model_flagged_ratios.csv

@@ -0,0 +1,2 @@
+model_name,flagged/all
+Meta-Llama-3-8B,0.31

eval_harmfulness/eval_results/llama3_eval/seed_123/flagged_ratio.csv

@@ -0,0 +1,2 @@
+model_name,flagged/all
+Meta-Llama-3-8B,0.32285714285714284

eval_harmfulness/eval_results/llama3_eval/seed_42/flagged_ratio.csv

@@ -0,0 +1,2 @@
+model_name,flagged/all
+Meta-Llama-3-8B,0.32857142857142857

eval_harmfulness/eval_results/llama3_eval/seed_456/flagged_ratio.csv

@@ -0,0 +1,2 @@
+model_name,flagged/all
+Meta-Llama-3-8B,0.31857142857142856

eval_harmfulness/eval_results/test_llama2/flagged_ratio.csv

@@ -0,0 +1,2 @@
+model_name,flagged/all
+Llama-2-7b-hf,0.31142857142857144

eval_harmfulness/eval_results/test_olmo_7b/flagged_ratio.csv

@@ -0,0 +1,2 @@
+model_name,flagged/all
+OLMo-7B-hf,0.28714285714285714

eval_harmfulness/evaluate_outputs.py

@@ -28,6 +28,8 @@
 import pandas as pd
 from evaluation_scripts.moderation import QAModeration
 from evaluation_scripts.parse_args import parse_arguments
+from utils import reproducibility
+import torch
 
 
 def calculate_flagged_proportion_and_agreement(data: dict) -> dict:
@@ -77,6 +79,9 @@ def plot_metrics(metrics: list[dict], output_dir: str) -> None:
 def main() -> None:
     args = parse_arguments()
 
+    # seeding
+    seed = reproducibility(args.seed)
+
     if not args.use_existing_evaluation:
         assert (
             args.eval_dataset is not None
@@ -106,6 +111,7 @@ def main() -> None:
             model_max_length=args.max_length,
             device_map=args.device,
         )
+
         predictions = model.predict(
             question=prompts,
             answer=responses,

eval_harmfulness/evaluation_scripts/moderation.py

@@ -141,6 +141,8 @@ def from_pretrained(
             model_kwargs["problem_type"] = problem_type
         if device_map is not None:
             model_kwargs["device_map"] = device_map
+        # temp modification on running at half precision
+        model_kwargs["torch_dtype"] = torch.bfloat16
 
         model = AutoModelForSequenceClassification.from_pretrained(
             model_name_or_path,

eval_harmfulness/evaluation_scripts/parse_args.py

@@ -46,4 +46,12 @@ def parse_arguments() -> argparse.Namespace:
         action="store_true",
         help="Uses already existing evaluation file at output_dir to just print and plot its contents!",
     )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="Seed for random number generators",
+    )
+
     return parser.parse_args()

eval_harmfulness/generation_scripts/parse_args.py

@@ -57,6 +57,13 @@ def parse_args() -> argparse.Namespace:
         help="Directory to which the output JSON is saved.",
     )
 
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="Seed for random number generators",
+    )
+
     args = parser.parse_args()
 
     return args