Upload 3 files
Browse files- config.json +4 -1
- decision_tree.pkl +1 -1
- modeling_decision_tree_reward_model.py +97 -0
config.json
CHANGED
@@ -1,7 +1,10 @@
|
|
1 |
{
|
2 |
"_name_or_path": "RLHFlow/Decision-Tree-Reward-Llama-3.1-8B",
|
|
|
|
|
|
|
3 |
"architectures": [
|
4 |
-
"
|
5 |
],
|
6 |
"attention_bias": false,
|
7 |
"attention_dropout": 0.0,
|
|
|
1 |
{
|
2 |
"_name_or_path": "RLHFlow/Decision-Tree-Reward-Llama-3.1-8B",
|
3 |
+
"auto_map": {
|
4 |
+
"AutoModelForSequenceClassification": "modeling_decision_tree_reward_model.LlamaForDecisionTreeRewardModel"
|
5 |
+
},
|
6 |
"architectures": [
|
7 |
+
"LlamaForDecisionTreeRewardModel"
|
8 |
],
|
9 |
"attention_bias": false,
|
10 |
"attention_dropout": 0.0,
|
decision_tree.pkl
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
size 2388
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:83f0139429fff38e775af9a281ba5600a46ff852967f6c310667e61710b5bf40
|
3 |
size 2388
|
modeling_decision_tree_reward_model.py
ADDED
@@ -0,0 +1,97 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
from transformers.models.llama.modeling_llama import LlamaForSequenceClassification
|
4 |
+
from sklearn.tree import DecisionTreeClassifier
|
5 |
+
import os
|
6 |
+
import pickle
|
7 |
+
import json
|
8 |
+
from huggingface_hub import hf_hub_download
|
9 |
+
from typing import List, Dict, Union
|
10 |
+
import numpy as np
|
11 |
+
|
12 |
+
def convert_to_chat_format(prompt, response=None):
|
13 |
+
if "<extra_id_1>" in prompt:
|
14 |
+
"""
|
15 |
+
Handling HelpSteer2 prompts which may contain multi-turn conversations with the special token <extra_id_1>
|
16 |
+
"""
|
17 |
+
turns = prompt.split("<extra_id_1>")
|
18 |
+
conversation = []
|
19 |
+
conversation.append({
|
20 |
+
"role": "user",
|
21 |
+
"content": turns[0]
|
22 |
+
})
|
23 |
+
|
24 |
+
for i in range(1, len(turns)):
|
25 |
+
parts = turns[i].split("\n", 1)
|
26 |
+
role = parts[0]
|
27 |
+
content = parts[1]
|
28 |
+
conversation.append({
|
29 |
+
"role": "assistant" if role == "Assistant" else "user",
|
30 |
+
"content": content
|
31 |
+
})
|
32 |
+
else:
|
33 |
+
conversation = [{"role": "user", "content": prompt}]
|
34 |
+
if response is not None:
|
35 |
+
conversation.append({"role": "assistant", "content": response})
|
36 |
+
return conversation
|
37 |
+
|
38 |
+
def process_conversation(conversation):
|
39 |
+
for message in conversation:
|
40 |
+
message["content"] = message["content"].rstrip('\n')
|
41 |
+
return conversation
|
42 |
+
|
43 |
+
class LlamaForDecisionTreeRewardModel(LlamaForSequenceClassification):
|
44 |
+
def __init__(self, config):
|
45 |
+
super().__init__(config)
|
46 |
+
self.score = nn.Linear(config.hidden_size, self.num_labels, bias=True)
|
47 |
+
# Initialize the decision tree
|
48 |
+
self.tree = None
|
49 |
+
# Define the default attributes (from HelpSteer2)
|
50 |
+
self.attributes = ['helpfulness', 'correctness', 'coherence', 'complexity', 'verbosity']
|
51 |
+
print("Initialized LlamaForDecisionTreeRewardModel")
|
52 |
+
|
53 |
+
def load_decision_tree(self, repo_id, filename="decision_tree.pkl"):
|
54 |
+
# Load the tree from the model's directory
|
55 |
+
with open(hf_hub_download(repo_id=repo_id, filename=filename), "rb") as f:
|
56 |
+
self.tree = pickle.load(f)
|
57 |
+
assert isinstance(self.tree, DecisionTreeClassifier), f"The tree is not a DecisionTreeClassifier. It is a {type(self.tree)}"
|
58 |
+
with open(hf_hub_download(repo_id=repo_id, filename="config.json"), "r") as f:
|
59 |
+
config = json.load(f)
|
60 |
+
label2id_map = config["label2id"]
|
61 |
+
# Sort labels and ids by ids
|
62 |
+
labels, ids = zip(*sorted(label2id_map.items(), key=lambda x: x[1]))
|
63 |
+
labels = list(labels)
|
64 |
+
self.attributes = labels
|
65 |
+
|
66 |
+
@torch.no_grad()
|
67 |
+
def compare(self, prompt: Union[str, List[Dict[str, str]]], response_1: str, response_2: str, tokenizer, device):
|
68 |
+
"""
|
69 |
+
Compare two inputs and return the difference in scores
|
70 |
+
"""
|
71 |
+
assert self.tree is not None, "The decision tree is not loaded. Please call load_decision_tree(repo_id, filename) first."
|
72 |
+
if isinstance(prompt, str):
|
73 |
+
conversation = convert_to_chat_format(prompt)
|
74 |
+
elif isinstance(prompt, list):
|
75 |
+
conversation = prompt
|
76 |
+
else:
|
77 |
+
raise ValueError(f"The prompt must be a string or a list of dictionaries, but got {type(prompt)}")
|
78 |
+
assert isinstance(conversation, list), "The conversation must be a list of dictionaries"
|
79 |
+
assert len(conversation) >= 1, "The conversation must have at least one message (as prompt)"
|
80 |
+
assert conversation[-1]["role"] == "user", "The last message in the conversation must be from the user"
|
81 |
+
conversation_1 = conversation + [{"role": "assistant", "content": response_1}]
|
82 |
+
conversation_2 = conversation + [{"role": "assistant", "content": response_2}]
|
83 |
+
conversation_1 = process_conversation(conversation_1)
|
84 |
+
conversation_2 = process_conversation(conversation_2)
|
85 |
+
|
86 |
+
conv_tokenized_1 = tokenizer.apply_chat_template(conversation_1, tokenize=True, return_tensors="pt").to(device)
|
87 |
+
conv_tokenized_2 = tokenizer.apply_chat_template(conversation_2, tokenize=True, return_tensors="pt").to(device)
|
88 |
+
embedding_1 = self.forward(conv_tokenized_1, output_hidden_states=True).hidden_states[-1][:,-1].float().cpu().numpy()
|
89 |
+
embedding_2 = self.forward(conv_tokenized_2, output_hidden_states=True).hidden_states[-1][:,-1].float().cpu().numpy()
|
90 |
+
weight = self.score.weight.float().cpu().numpy()
|
91 |
+
bias = self.score.bias.float().cpu().numpy()
|
92 |
+
rewards_1 = embedding_1 @ weight.T + bias
|
93 |
+
rewards_2 = embedding_2 @ weight.T + bias
|
94 |
+
rewards_diff = rewards_2 - rewards_1
|
95 |
+
return {
|
96 |
+
"preference": self.tree.predict(rewards_diff)[0], "rewards": np.stack([rewards_1, rewards_2]),
|
97 |
+
"attributes": self.attributes}
|