Merge branch 'gptj' of github.com:nomic-ai/gpt4all into gptj

GPTJ.md

@@ -0,0 +1,17 @@
+# Inference on Training Data
+
+
+## Run Inference
+
+```bash
+torchrun --master_port=29085 --nproc-per-node 8 inference.py --config=configs/inference/gptj.yaml
+```
+
+
+## Visualizations
+
+```bash
+python build_map.py
+```
+ 
+will build a map in `Atlas`, one using the internal clustering algorithm provided by Nomic and one using the embeddings generated by the finetuned model.

configs/eval/generate.yaml

@@ -1,15 +1,5 @@
 # model/tokenizer
-model_name: # update with llama 7b 
+model_name: # update with llama model name
-tokenizer_name: # update with llama 7b
+tokenizer_name: # update with llama model name
 lora: true
 lora_path: "nomic-ai/gpt4all-lora"
-
-max_new_tokens: 512
-temperature: 0.001
-prompt: | 
-  #this code prints a string reversed
-  my_string = "hello how are you"
-  print(len(my_string))
-
-
-  My code above does not work. Can you help me?

configs/eval/generate_baseline.yaml

@@ -2,16 +2,4 @@
 model_name: # update with llama model name
 tokenizer_name: # update with llama model name
 lora: true
-lora_path: "tloen/alpaca-lora-7b"
+lora_path: "tloen/alpaca-lora-7b"
-
-
-
-max_new_tokens: 512
-temperature: 0.001
-prompt: | 
-  #this code prints a string reversed
-  my_string = "hello how are you"
-  print(len(my_string))
-
-
-  My code above does not work. Can you help me?

configs/eval/generate_full.yaml

@@ -1,14 +0,0 @@
-# model/tokenizer
-model_name: # update
-tokenizer_name: # update
-lora_path: "no-lora"
-
-max_new_tokens: 512
-temperature: 0.001
-prompt: | 
-  #this code prints a string reversed
-  my_string = "hello how are you"
-  print(len(my_string))
-
-
-  My code above does not work. Can you help me?

configs/eval/generate_gpt4all_gptj.yaml

@@ -0,0 +1,4 @@
+# model/tokenizer
+model_name: "nomic-ai/gpt4all-warmup-lr-epoch_1"
+tokenizer_name: "EleutherAI/gpt-j-6b"
+lora: false

configs/eval/generate_gpt4all_gptj_lora.yaml

@@ -0,0 +1,5 @@
+# model/tokenizer
+model_name: "EleutherAI/gpt-j-6b"
+tokenizer_name: "EleutherAI/gpt-j-6B"
+lora: true
+lora_path: "nomic-ai/gpt4all-gptj-lora-epoch_1"

configs/eval/generate_large_2.yaml

@@ -1,15 +0,0 @@
-# model/tokenizer
-model_name: # update
-tokenizer_name: # update
-lora: true
-lora_path: # update
-
-max_new_tokens: 512
-temperature: 0.001
-prompt: | 
-  #this code prints a string reversed
-  my_string = "hello how are you"
-  print(len(my_string))
-
-
-  My code above does not work. Can you help me?

configs/eval/generate_large_3.yaml

@@ -1,15 +0,0 @@
-# model/tokenizer
-model_name: # update
-tokenizer_name: # update
-lora: true
-lora_path: # update
-
-max_new_tokens: 512
-temperature: 0.001
-prompt: | 
-  #this code prints a string reversed
-  my_string = "hello how are you"
-  print(len(my_string))
-
-
-  My code above does not work. Can you help me?

configs/inference/gptj.yaml

@@ -1,11 +1,11 @@
 # model/tokenizer
-model_name: "nomic-ai/gpt4all-gptj-multinode-deepspeed-finetuned-epoch_0"
+model_name: "nomic-ai/gpt4all-warmup-lr-epoch_1"
 tokenizer_name: "EleutherAI/gpt-j-6B"
 
 # dataset
 streaming: false
 num_proc: 64
-dataset_path: "data_multiplus" 
+dataset_path: "nomic-ai/turbo-500k-multi" 
 max_length: 1024
 batch_size: 32 
 

configs/train/finetune_gptj.yaml

@@ -2,14 +2,14 @@
 model_name: "EleutherAI/gpt-j-6B"
 tokenizer_name: "EleutherAI/gpt-j-6B"
 gradient_checkpointing: true
-save_name: "nomic-ai/gpt4all-mosaic"
+save_name: "nomic-ai/gpt4all-warmup-lr"
 
 # dataset
 streaming: false
 num_proc: 64
 dataset_path: "nomic-ai/turbo-500k-multi"
 max_length: 1024
-batch_size: 8
+batch_size: 32
 
 # train dynamics
 lr: 2.0e-5

eval_figures.py

@@ -6,18 +6,20 @@ from matplotlib import pyplot as plt
 plt.figure()
 for fpath in glob.glob('./eval_data/*.pkl'):
     parts = fpath.split('__')
-    model_name = parts[1].replace('model-', '').replace('.pkl', '')
+    model_name = "-".join(fpath.replace(".pkl", "").split("_")[2:])
-    lora_name = parts[2].replace('lora-', '').replace('.pkl', '')
     with open(fpath, 'rb') as f:
         data = pickle.load(f)
         perplexities = data['perplexities']
         perplexities = np.nan_to_num(perplexities, 100)
         perplexities = np.clip(perplexities, 0, 100)
-        if 'nomic' in fpath:
+        if 'alpaca' not in fpath:
-            label = 'GPT4all-lora'
+            identifier = model_name = "-".join(fpath.replace(".pkl", "").split("eval__model-")[1:]) 
+            label = 'GPT4all-'
+            label += identifier
+            
         else:
             label = 'alpaca-lora'
-        plt.hist(perplexities, label=label, alpha=.5)
+        plt.hist(perplexities, label=label, alpha=.5, bins=50)
 
 plt.xlabel('Perplexity')
 plt.ylabel('Frequency')

eval_self_instruct.py

@@ -49,28 +49,6 @@ def eval_example(model, tokenizer, example, config):
     input = tokenizer(prompt, return_tensors="pt")
     input = {k: v.to(model.device) for k, v in input.items()}
 
-    continuations = []
-    tokenized_continuations = []
-    trajectories = []
-    for i in range(1):
-        with torch.no_grad():
-            outputs = model.generate(input_ids=input['input_ids'],
-                                     max_new_tokens=config["max_new_tokens"],
-                                     min_new_tokens=5,
-                                     temperature=config["temperature"],
-                                     repetition_penalty=1.0,
-                                     do_sample=True)
-            decoded = tokenizer.decode(outputs[0], skip_special_tokens=True).strip()
-
-            y = model(input_ids=outputs)
-        trajectory = y.hidden_states[0].detach().cpu().numpy()[0]
-        trajectory = trajectory / np.linalg.norm(trajectory, axis=1, keepdims=True)
-        trajectory = np.cumsum(trajectory, axis=0) / np.arange(1, trajectory.shape[0]+1).reshape(-1, 1)
-
-        trajectories.append(trajectory)
-        continuations.append(decoded)
-        tokenized_continuations.append(tokenizer.tokenize(decoded))
-
     #compute the ground truth perplexity
     gt_input = tokenizer(gt, return_tensors="pt")
     gt_input = {k: v.to(model.device) for k, v in gt_input.items()}
@@ -101,30 +79,23 @@ def eval_example(model, tokenizer, example, config):
 
     print(prompt)
     print(80*'-')
-    for continuation in continuations:
+   
-        print(continuation)
-        print(80*'-')
 
-    return ppl, trajectories, continuations, tokenized_continuations
+    return ppl
 
 def do_eval(config):
     eval_data = read_jsonl_file('eval_data/user_oriented_instructions.jsonl')
     model, tokenizer = setup_model(config)
-    all_trajectories = []
     all_perplexities = []
-    all_continuations = []
-    all_tokenized_continuations = []
     for example in tqdm(eval_data):
-        gt_perplexity, trajectories, continuations, tokenized_continuations = eval_example(model, tokenizer, example, config)
+        gt_perplexity = eval_example(model, tokenizer, example, config)
-        all_trajectories.append(trajectories)
         all_perplexities.append(gt_perplexity)
-        all_continuations.append(continuations)
 
-    with open('eval_data/eval__model-{}__lora-{}.pkl'.format(config['model_name'].replace('/', '_'), config['lora_path'].replace('/', '_')), 'wb') as f:
+        
-        r = {'trajectories': all_trajectories,
+    name = f"eval_data/eval__model-{config['model_name'].replace('/', '_')}{'__lora-' + config['lora_path'].replace('/', '_') if config['lora'] else ''}.pkl"
-             'perplexities': all_perplexities,
+
-             'continuations': all_continuations,
+    with open(name, 'wb') as f:
-             'tokenized_continuations': all_tokenized_continuations}
+        r = {'perplexities': all_perplexities}
         pickle.dump(r, f)
 
 

requirements.txt

@@ -11,4 +11,5 @@ deepspeed
 sentencepiece
 jsonlines
 nomic
-scikit-learn
+scikit-learn
+matplotlib