File size: 2,316 Bytes
d4f8732
 
 
 
bf42bfb
 
d4f8732
 
540ce3f
d4f8732
540ce3f
d4f8732
540ce3f
d4f8732
 
 
 
540ce3f
 
 
 
 
 
 
 
d4f8732
 
 
 
540ce3f
 
 
 
d4f8732
 
 
540ce3f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d4f8732
540ce3f
d4f8732
 
 
 
 
540ce3f
 
 
 
 
 
 
 
 
 
d4f8732
540ce3f
d4f8732
540ce3f
d4f8732
540ce3f
d4f8732
 
 
540ce3f
 
 
 
 
 
d4f8732
 
 
540ce3f
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
---
library_name: transformers
tags:
- llama-factory
base_model:
- mistralai/Mistral-7B-Instruct-v0.3
---

## Summary

This model is a fine-tuned version of the Mistral-7B-Instruct-v0.3 optimised for answering questions in the domain of forensic investigations. The model has been trained using a specialised dataset titled **Advanced_Forensic_Investigations_Knowledge_Library_v1**, which consists of approximately 100 domain-specific question-answer pairs. The objective is to support advanced forensic investigative reasoning, rapid knowledge retrieval, and high-precision forensic domain assistance.

---

## Model Details

### Model Description
    
- **Model type:** Instruction-following Language Model (LoRA-based fine-tuning)
    
- **Language(s):** English
    
- **Fine-tuned from model:** Mistral-7B-Instruct-v0.3
    
---
## Training Details

### Training Data

- **Dataset:** `Advanced_Forensic_Investigations_Knowledge_Library_v1`
    
- **Data size:** ~100 high-quality, domain-specific QA pairs
    

### Training Procedure

#### Preprocessing

- Template: `mistral`
    
- Token truncation/cutoff: 2048
    
- No vocab resizing or prompt packing
    

#### Hyperparameters

- **Finetuning type:** LoRA
    
- **Precision:** `bf16`
    
- **LoRA rank:** 16
    
- **LoRA alpha:** 32
    
- **Batch size:** 4
    
- **Gradient accumulation:** 8
    
- **Learning rate:** `3e-4`
    
- **Epochs:** 35
    
- **LR scheduler:** cosine
    
- **Quantisation:** 4-bit (bitsandbytes)
    
- **Cutoff length:** 2048
    

#### Compute
    
- **Training time:** close to 30 minutes
    
- **Framework:** LLaMA-Factory

---

## Evaluation

### Testing Data, Factors & Metrics

- **Metrics Used:** BLEU-4, ROUGE-1, ROUGE-2
    
- **Results:**
    
    - BLEU-4: 100%
        
    - ROUGE-1: 100%
        
    - ROUGE-2: 100%
        

These scores reflect perfect overlap with reference answers within the scope of the evaluation dataset.

---

## Technical Specifications

### Model Architecture and Objective

- **Base:** Transformer (Mistral-7B architecture)
    
- **Fine-tuning method:** LoRA
    
- **Objective:** Instruction-following with forensic legal knowledge adaptation
    

### Compute Infrastructure

- **Hardware:** 2xL40s
    
- **Software:** LLaMA-Factory, PyTorch, Transformers, bitsandbytes