Liam Wilbur commited on
Commit
6755bcd
·
1 Parent(s): 95d91c5

Added all-miniLM-l6-v2 transformer fine-tuned on synonyms

Browse files
Files changed (1) hide show
  1. README.md +6 -16
README.md CHANGED
@@ -20,17 +20,6 @@ A fine-tuned SentenceTransformer model specialized for finding synonyms and rela
20
  This model is fine-tuned from `sentence-transformers/all-MiniLM-L6-v2` specifically for automotive parts synonym detection. It can identify when different part names refer to the same or similar components.
21
 
22
  **Base Model:** `sentence-transformers/all-MiniLM-L6-v2`
23
- **Output Dimensions:** 384
24
- **Max Sequence Length:** 256 tokens
25
-
26
- ## Intended Use
27
-
28
- - **Primary:** Finding synonymous automotive parts and service terms
29
- - **Secondary:** Semantic search for automotive repair descriptions
30
- - **Examples:**
31
- - "Drive Motor" ↔ "gear motor"
32
- - "install new Actuator Switch" ↔ "replace fan switch"
33
- - "TXV Valve" ↔ "Expansion Valve"
34
 
35
  ## Training Details
36
 
@@ -41,16 +30,17 @@ This model is fine-tuned from `sentence-transformers/all-MiniLM-L6-v2` specifica
41
 
42
  **Loss Function:** OnlineContrastiveLoss with varying margins (0.6 → 0.4 → 0.4)
43
  **Training Data:** Automotive parts synonym/antonym pairs with contextual repair descriptions
44
- **LoRA:** Used for parameter-efficient fine-tuning (r=16, alpha=16)
45
 
46
  ## Performance
47
 
48
- Evaluated on STS-B and MTB datasets:
 
 
 
 
49
  - **STS-B Spearman:** 0.867
50
  - **MTB Spearman:** 0.724
51
 
52
  ## Limitations
53
-
54
  - Optimized specifically for automotive parts and repair terminology
55
- - May not perform well on general-domain text
56
- - Best results on short phrases (3-10 tokens) typical of part names
 
20
  This model is fine-tuned from `sentence-transformers/all-MiniLM-L6-v2` specifically for automotive parts synonym detection. It can identify when different part names refer to the same or similar components.
21
 
22
  **Base Model:** `sentence-transformers/all-MiniLM-L6-v2`
 
 
 
 
 
 
 
 
 
 
 
23
 
24
  ## Training Details
25
 
 
30
 
31
  **Loss Function:** OnlineContrastiveLoss with varying margins (0.6 → 0.4 → 0.4)
32
  **Training Data:** Automotive parts synonym/antonym pairs with contextual repair descriptions
 
33
 
34
  ## Performance
35
 
36
+ Evaluated on 15-state synonyms phrases, deepseek general queries, STS-B and MTB datasets:
37
+ - **Top-200 Synonym Recall** 15251/15330 (99.5%)
38
+ - **MRR@200** 0.6454
39
+ - **Top-200 Deepseek Recall** 140633/142325 (98.8%)
40
+ - **MRR@200** 0.4401
41
  - **STS-B Spearman:** 0.867
42
  - **MTB Spearman:** 0.724
43
 
44
  ## Limitations
 
45
  - Optimized specifically for automotive parts and repair terminology
46
+ - May not perform well on general-domain text