doberst commited on
Commit
cccaae8
1 Parent(s): 36efaaa

Upload README.md

Browse files
Files changed (1) hide show
  1. README.md +122 -0
README.md CHANGED
@@ -1,3 +1,125 @@
1
  ---
2
  license: apache-2.0
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: apache-2.0
3
  ---
4
+
5
+ # Model Card for Model ID
6
+
7
+ <!-- Provide a quick summary of what the model is/does. -->
8
+
9
+ dragon-mistral-7b-v0 part of the dRAGon ("Delivering RAG On ...") model series, RAG-instruct trained on top of a Mistral-7B base model.
10
+
11
+ DRAGON models are fine-tuned with high-quality custom instruct datasets, designed for production quality use in RAG scenarios.
12
+
13
+
14
+ ### Benchmark Tests
15
+
16
+ Evaluated against the benchmark test: [RAG-Instruct-Benchmark-Tester](https://www.huggingface.co/datasets/llmware/rag_instruct_benchmark_tester)
17
+ Average of 2 Test Runs with 1 point for correct answer, 0.5 point for partial correct or blank / NF, 0.0 points for incorrect, and -1 points for hallucinations.
18
+
19
+ --**Accuracy Score**: **94.75** correct out of 100
20
+ --Not Found Classification: 95.0%
21
+ --Boolean: 94.0%
22
+ --Math/Logic: 77.5%
23
+ --Complex Questions (1-5): 4 (Low-Medium)
24
+ --Summarization Quality (1-5): 4 (Coherent, extractive)
25
+ --Hallucinations: No hallucinations observed in test runs.
26
+
27
+ For test run results (and good indicator of target use cases), please see the files ("core_rag_test" and "answer_sheet" in this repo).
28
+
29
+ ### Model Description
30
+
31
+ <!-- Provide a longer summary of what this model is. -->
32
+
33
+ - **Developed by:** llmware
34
+ - **Model type:** Mistral-7B
35
+ - **Language(s) (NLP):** English
36
+ - **License:** Apache 2.0
37
+ - **Finetuned from model:** Mistral-7B-Base
38
+
39
+ ## Uses
40
+
41
+ <!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
42
+
43
+ The intended use of DRAGON models is two-fold:
44
+
45
+ 1. Provide high-quality RAG-Instruct models designed for fact-based, no "hallucination" question-answering in connection with an enterprise RAG workflow.
46
+
47
+ 2. DRAGON models are fine-tuned on top of leading base foundation models, generally in the 6-7B+ range, and purposefully rolled-out across multiple base models to provide choices and "drop-in" replacements for RAG specific use cases.
48
+
49
+ 3. DRAGON models were trained on the same principles as the BLING models, so generally, it should be easy to "upgrade" from a BLING model in testing to a DRAGON model in production.
50
+
51
+
52
+ ### Direct Use
53
+
54
+ <!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
55
+
56
+ DRAGON is designed for enterprise automation use cases, especially in knowledge-intensive industries, such as financial services,
57
+ legal and regulatory industries with complex information sources.
58
+
59
+ DRAGON models have been trained for common RAG scenarios, specifically: question-answering, key-value extraction, and basic summarization as the core instruction types
60
+ without the need for a lot of complex instruction verbiage - provide a text passage context, ask questions, and get clear fact-based responses.
61
+
62
+
63
+ ## Bias, Risks, and Limitations
64
+
65
+ <!-- This section is meant to convey both technical and sociotechnical limitations. -->
66
+
67
+ Any model can provide inaccurate or incomplete information, and should be used in conjunction with appropriate safeguards and fact-checking mechanisms.
68
+
69
+
70
+ ## How to Get Started with the Model
71
+
72
+ The fastest way to get started with dRAGon is through direct import in transformers:
73
+
74
+ from transformers import AutoTokenizer, AutoModelForCausalLM
75
+ tokenizer = AutoTokenizer.from_pretrained("dragon-mistral-7b-v0")
76
+ model = AutoModelForCausalLM.from_pretrained("dragon-mistral-7b-v0")
77
+
78
+ Please refer to the generation_test .py files in the Files repository, which includes 200 samples and script to test the model. The **generation_test_llmware_script.py** includes built-in llmware capabilities for fact-checking, as well as easy integration with document parsing and actual retrieval to swap out the test set for RAG workflow consisting of business documents.
79
+
80
+ The dRAGon model was fine-tuned with a simple "\<human> and \<bot> wrapper", so to get the best results, wrap inference entries as:
81
+
82
+ full_prompt = "<human>: " + my_prompt + "\n" + "<bot>:"
83
+
84
+ The BLING model was fine-tuned with closed-context samples, which assume generally that the prompt consists of two sub-parts:
85
+
86
+ 1. Text Passage Context, and
87
+ 2. Specific question or instruction based on the text passage
88
+
89
+ To get the best results, package "my_prompt" as follows:
90
+
91
+ my_prompt = {{text_passage}} + "\n" + {{question/instruction}}
92
+
93
+
94
+ If you are using a HuggingFace generation script:
95
+
96
+ # prepare prompt packaging used in fine-tuning process
97
+ new_prompt = "<human>: " + entries["context"] + "\n" + entries["query"] + "\n" + "<bot>:"
98
+
99
+ inputs = tokenizer(new_prompt, return_tensors="pt")
100
+ start_of_output = len(inputs.input_ids[0])
101
+
102
+ # temperature: set at 0.3 for consistency of output
103
+ # max_new_tokens: set at 100 - may prematurely stop a few of the summaries
104
+
105
+ outputs = model.generate(
106
+ inputs.input_ids.to(device),
107
+ eos_token_id=tokenizer.eos_token_id,
108
+ pad_token_id=tokenizer.eos_token_id,
109
+ do_sample=True,
110
+ temperature=0.3,
111
+ max_new_tokens=100,
112
+ )
113
+
114
+ output_only = tokenizer.decode(outputs[0][start_of_output:],skip_special_tokens=True)
115
+
116
+ # note: due to artifact of the fine-tuning, use this post-processing with HF generation
117
+
118
+ eot = output_only.find("<|endoftext|>")
119
+ if eot > -1:
120
+ output_only = output_only[:eot]
121
+
122
+
123
+ ## Model Card Contact
124
+
125
+ Darren Oberst & llmware team