Shivenys commited on
Commit
1979f80
·
verified ·
1 Parent(s): 5ecc74e

Initial upload of Phi-3 LoRA adapter

Browse files
README.md CHANGED
@@ -1,3 +1,163 @@
1
  ---
2
  license: apache-2.0
 
 
 
 
 
 
 
 
 
 
 
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: apache-2.0
3
+ base_model: microsoft/Phi-3-mini-4k-instruct
4
+ library_name: peft
5
+ tags:
6
+ - phi-3
7
+ - lora
8
+ - qlora
9
+ - peft
10
+ - transformers
11
+ - instruction-tuning
12
+ - fine-tuning
13
+ - pytorch
14
  ---
15
+
16
+ # Phi-3 Mini 4K Instruct - LoRA Fine-Tuned Model
17
+
18
+ This repository contains a **QLoRA (4-bit LoRA)** adapter fine-tuned from **Microsoft Phi-3 Mini 4K Instruct** using the Hugging Face Transformers, PEFT, and TRL libraries.
19
+
20
+ > **Note:** This model was created as a learning project to understand the complete LLM fine-tuning workflow.
21
+
22
+ ---
23
+
24
+ ## Base Model
25
+
26
+ - **Base Model:** `microsoft/Phi-3-mini-4k-instruct`
27
+ - **Fine-Tuning Method:** QLoRA (4-bit Quantization + LoRA)
28
+ - **Framework:** Hugging Face Transformers
29
+ - **PEFT Library:** PEFT
30
+ - **Trainer:** TRL SFTTrainer
31
+
32
+ ---
33
+
34
+ ## Model Details
35
+
36
+ | Parameter | Value |
37
+ |-----------|-------|
38
+ | Base Model | Phi-3 Mini 4K Instruct |
39
+ | Fine-Tuning Method | QLoRA |
40
+ | LoRA Rank (r) | 8 |
41
+ | LoRA Alpha | 16 |
42
+ | LoRA Dropout | 0.05 |
43
+ | Task | Causal Language Modeling |
44
+ | Quantization | 4-bit NF4 |
45
+ | Compute Type | FP16 |
46
+
47
+ ---
48
+
49
+ ## Training Configuration
50
+
51
+ - Epochs: **5**
52
+ - Batch Size: **1**
53
+ - Learning Rate: **2e-4**
54
+ - Max Sequence Length: **256**
55
+ - Optimizer: AdamW (default TRL optimizer)
56
+
57
+ ---
58
+
59
+ ## Training Dataset
60
+
61
+ The adapter was fine-tuned on a small custom instruction-response dataset containing sample examples such as:
62
+
63
+ - Who created Python?
64
+ - What is AI?
65
+ - What is LoRA?
66
+ - What is the capital of France?
67
+
68
+ This dataset was intentionally kept small for educational purposes to demonstrate the end-to-end fine-tuning process.
69
+
70
+ ---
71
+
72
+ ## Usage
73
+
74
+ Install the required libraries:
75
+
76
+ ```bash
77
+ pip install transformers peft bitsandbytes accelerate torch
78
+ ```
79
+
80
+ Load the adapter:
81
+
82
+ ```python
83
+ import torch
84
+ from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
85
+ from peft import PeftModel
86
+
87
+ BASE_MODEL = "microsoft/Phi-3-mini-4k-instruct"
88
+
89
+ bnb_config = BitsAndBytesConfig(
90
+ load_in_4bit=True,
91
+ bnb_4bit_quant_type="nf4",
92
+ bnb_4bit_compute_dtype=torch.float16,
93
+ bnb_4bit_use_double_quant=True,
94
+ )
95
+
96
+ tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
97
+
98
+ base_model = AutoModelForCausalLM.from_pretrained(
99
+ BASE_MODEL,
100
+ quantization_config=bnb_config,
101
+ device_map="auto",
102
+ )
103
+
104
+ model = PeftModel.from_pretrained(
105
+ base_model,
106
+ "YOUR_USERNAME/YOUR_REPOSITORY_NAME"
107
+ )
108
+ ```
109
+
110
+ Generate text:
111
+
112
+ ```python
113
+ prompt = """### Instruction:
114
+ Who created Python?
115
+
116
+ ### Response:
117
+ """
118
+
119
+ inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
120
+
121
+ outputs = model.generate(
122
+ **inputs,
123
+ max_new_tokens=100,
124
+ )
125
+
126
+ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
127
+ ```
128
+
129
+ ---
130
+
131
+ ## Repository Structure
132
+
133
+ ```
134
+ adapter_config.json
135
+ adapter_model.safetensors
136
+ README.md
137
+ ```
138
+
139
+ ---
140
+
141
+ ## Limitations
142
+
143
+ - This adapter was trained on a very small dataset.
144
+ - It is intended for educational and demonstration purposes.
145
+ - It should not be considered a production-ready fine-tuned model.
146
+
147
+ ---
148
+
149
+ ## Acknowledgements
150
+
151
+ - Microsoft for the Phi-3 Mini 4K Instruct model
152
+ - Hugging Face Transformers
153
+ - Hugging Face TRL
154
+ - PEFT
155
+ - BitsAndBytes
156
+
157
+ ---
158
+
159
+ ## License
160
+
161
+ This repository contains only the LoRA adapter.
162
+
163
+ Please follow the license terms of the original Phi-3 Mini 4K Instruct model when using the base model.
adapter_config.json ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "alora_invocation_tokens": null,
3
+ "alpha_pattern": {},
4
+ "arrow_config": null,
5
+ "auto_mapping": null,
6
+ "base_model_name_or_path": "microsoft/Phi-3-mini-4k-instruct",
7
+ "bias": "none",
8
+ "corda_config": null,
9
+ "ensure_weight_tying": false,
10
+ "eva_config": null,
11
+ "exclude_modules": null,
12
+ "fan_in_fan_out": false,
13
+ "inference_mode": true,
14
+ "init_lora_weights": true,
15
+ "layer_replication": null,
16
+ "layers_pattern": null,
17
+ "layers_to_transform": null,
18
+ "loftq_config": {},
19
+ "lora_alpha": 16,
20
+ "lora_bias": false,
21
+ "lora_dropout": 0.05,
22
+ "lora_ga_config": null,
23
+ "megatron_config": null,
24
+ "megatron_core": "megatron.core",
25
+ "modules_to_save": null,
26
+ "peft_type": "LORA",
27
+ "peft_version": "0.19.1",
28
+ "qalora_group_size": 16,
29
+ "r": 8,
30
+ "rank_pattern": {},
31
+ "revision": null,
32
+ "target_modules": [
33
+ "qkv_proj",
34
+ "o_proj"
35
+ ],
36
+ "target_parameters": null,
37
+ "task_type": "CAUSAL_LM",
38
+ "trainable_token_indices": null,
39
+ "use_bdlora": null,
40
+ "use_dora": false,
41
+ "use_qalora": false,
42
+ "use_rslora": false
43
+ }
adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:46410ac5be1667fdddfcc5aa28e64d5bc182b67bd10bc0497e7e342e00a9ffac
3
+ size 9454304
chat_template.jinja ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {% for message in messages %}{% if message['role'] == 'system' %}{{'<|system|>
2
+ ' + message['content'] + '<|end|>
3
+ '}}{% elif message['role'] == 'user' %}{{'<|user|>
4
+ ' + message['content'] + '<|end|>
5
+ '}}{% elif message['role'] == 'assistant' %}{{'<|assistant|>
6
+ ' + message['content'] + '<|end|>
7
+ '}}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|>
8
+ ' }}{% else %}{{ eos_token }}{% endif %}
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "tokenizers",
3
+ "bos_token": "<s>",
4
+ "clean_up_tokenization_spaces": false,
5
+ "eos_token": "<|endoftext|>",
6
+ "is_local": false,
7
+ "legacy": false,
8
+ "local_files_only": false,
9
+ "model_max_length": 4096,
10
+ "pad_token": "<|endoftext|>",
11
+ "padding_side": "left",
12
+ "sp_model_kwargs": {},
13
+ "tokenizer_class": "TokenizersBackend",
14
+ "unk_token": "<unk>",
15
+ "use_default_system_prompt": false
16
+ }
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9b83fc75ff391b29c0d29e57dd7512afa27fc43cb5af79c94338622050d9202f
3
+ size 5649