create readme.md and model card
Browse files
README.md
ADDED
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
datasets:
|
3 |
+
- Trelis/smollm-corpus-2percent
|
4 |
+
language:
|
5 |
+
- en
|
6 |
+
base_model:
|
7 |
+
- HuggingFaceTB/SmolLM-360M
|
8 |
+
tags:
|
9 |
+
- language_model
|
10 |
+
- pruned
|
11 |
+
- distilled
|
12 |
+
---
|
13 |
+
|
14 |
+
# Model Card for TrelisLM-80M
|
15 |
+
|
16 |
+
This model is a pruned and distilled version of SmolLM-360M, created for scientific curiosity.
|
17 |
+
|
18 |
+
## Model Details
|
19 |
+
|
20 |
+
### Model Description
|
21 |
+
|
22 |
+
- **Developed by:** Trelis Team
|
23 |
+
- **Model type:** Language Model
|
24 |
+
- **Language(s) (NLP):** English
|
25 |
+
- **License:** [More Information Needed]
|
26 |
+
- **Finetuned from model:** HuggingFaceTB/SmolLM-360M
|
27 |
+
|
28 |
+
TrelisLM-80M is a 80 million parameter language model derived from SmolLM-360M. It was created through a process of layer and width pruning, followed by distillation from SmolLM-360M-Instruct using Forward KL loss.
|
29 |
+
|
30 |
+
## Uses
|
31 |
+
|
32 |
+
### Direct Use
|
33 |
+
|
34 |
+
This model is primarily intended for scientific curiosity and research purposes. It can be used to explore the effects of model pruning and distillation on language model performance.
|
35 |
+
|
36 |
+
### Out-of-Scope Use
|
37 |
+
|
38 |
+
As this model is still not completely trained, it should not be used for any production or real-world applications at this stage.
|
39 |
+
|
40 |
+
## Bias, Risks, and Limitations
|
41 |
+
|
42 |
+
The model is still in the training process and may have unpredictable behaviors or biases. It should be used with caution and only for research purposes.
|
43 |
+
|
44 |
+
### Recommendations
|
45 |
+
|
46 |
+
Users should be aware that this model is a work in progress and its outputs should not be relied upon for any critical or sensitive tasks.
|
47 |
+
|
48 |
+
## Training Details
|
49 |
+
|
50 |
+
### Training Data
|
51 |
+
|
52 |
+
The model was distilled using the Trelis/smollm-corpus-2percent dataset.
|
53 |
+
|
54 |
+
### Training Procedure
|
55 |
+
|
56 |
+
The training procedure involved the following steps:
|
57 |
+
1. Layer pruning of SmolLM-360M
|
58 |
+
2. Width pruning of SmolLM-360M
|
59 |
+
3. Distillation from SmolLM-360M-Instruct using Forward KL loss
|
60 |
+
|
61 |
+
## Evaluation
|
62 |
+
|
63 |
+
Evaluation results are not yet available for this model.
|
64 |
+
|
65 |
+
## Model Examination
|
66 |
+
|
67 |
+
Further examination and interpretation of the model's behavior are needed.
|
68 |
+
|
69 |
+
## Environmental Impact
|
70 |
+
|
71 |
+
[More Information Needed]
|
72 |
+
|
73 |
+
## Technical Specifications
|
74 |
+
|
75 |
+
### Model Architecture and Objective
|
76 |
+
|
77 |
+
TrelisLM-80M is an 80 million parameter language model derived from SmolLM-360M through pruning and distillation.
|
78 |
+
|
79 |
+
### Compute Infrastructure
|
80 |
+
|
81 |
+
[More Information Needed]
|
82 |
+
|
83 |
+
## Model Card Contact
|
84 |
+
|
85 |
+
[More Information Needed]
|