RichardErkhov
commited on
Commit
•
3be9734
1
Parent(s):
b7f61f0
uploaded readme
Browse files
README.md
ADDED
@@ -0,0 +1,149 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Quantization made by Richard Erkhov.
|
2 |
+
|
3 |
+
[Github](https://github.com/RichardErkhov)
|
4 |
+
|
5 |
+
[Discord](https://discord.gg/pvy7H8DZMG)
|
6 |
+
|
7 |
+
[Request more models](https://github.com/RichardErkhov/quant_request)
|
8 |
+
|
9 |
+
|
10 |
+
Llama-3-13B-Instruct-ft - GGUF
|
11 |
+
- Model creator: https://huggingface.co/elinas/
|
12 |
+
- Original model: https://huggingface.co/elinas/Llama-3-13B-Instruct-ft/
|
13 |
+
|
14 |
+
|
15 |
+
| Name | Quant method | Size |
|
16 |
+
| ---- | ---- | ---- |
|
17 |
+
| [Llama-3-13B-Instruct-ft.Q2_K.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q2_K.gguf) | Q2_K | 4.68GB |
|
18 |
+
| [Llama-3-13B-Instruct-ft.IQ3_XS.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.IQ3_XS.gguf) | IQ3_XS | 5.18GB |
|
19 |
+
| [Llama-3-13B-Instruct-ft.IQ3_S.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.IQ3_S.gguf) | IQ3_S | 5.45GB |
|
20 |
+
| [Llama-3-13B-Instruct-ft.Q3_K_S.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q3_K_S.gguf) | Q3_K_S | 5.42GB |
|
21 |
+
| [Llama-3-13B-Instruct-ft.IQ3_M.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.IQ3_M.gguf) | IQ3_M | 5.61GB |
|
22 |
+
| [Llama-3-13B-Instruct-ft.Q3_K.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q3_K.gguf) | Q3_K | 5.98GB |
|
23 |
+
| [Llama-3-13B-Instruct-ft.Q3_K_M.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q3_K_M.gguf) | Q3_K_M | 5.98GB |
|
24 |
+
| [Llama-3-13B-Instruct-ft.Q3_K_L.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q3_K_L.gguf) | Q3_K_L | 6.47GB |
|
25 |
+
| [Llama-3-13B-Instruct-ft.IQ4_XS.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.IQ4_XS.gguf) | IQ4_XS | 6.69GB |
|
26 |
+
| [Llama-3-13B-Instruct-ft.Q4_0.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q4_0.gguf) | Q4_0 | 6.97GB |
|
27 |
+
| [Llama-3-13B-Instruct-ft.IQ4_NL.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.IQ4_NL.gguf) | IQ4_NL | 7.04GB |
|
28 |
+
| [Llama-3-13B-Instruct-ft.Q4_K_S.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q4_K_S.gguf) | Q4_K_S | 7.01GB |
|
29 |
+
| [Llama-3-13B-Instruct-ft.Q4_K.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q4_K.gguf) | Q4_K | 7.38GB |
|
30 |
+
| [Llama-3-13B-Instruct-ft.Q4_K_M.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q4_K_M.gguf) | Q4_K_M | 7.38GB |
|
31 |
+
| [Llama-3-13B-Instruct-ft.Q4_1.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q4_1.gguf) | Q4_1 | 7.7GB |
|
32 |
+
| [Llama-3-13B-Instruct-ft.Q5_0.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q5_0.gguf) | Q5_0 | 8.43GB |
|
33 |
+
| [Llama-3-13B-Instruct-ft.Q5_K_S.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q5_K_S.gguf) | Q5_K_S | 8.43GB |
|
34 |
+
| [Llama-3-13B-Instruct-ft.Q5_K.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q5_K.gguf) | Q5_K | 8.64GB |
|
35 |
+
| [Llama-3-13B-Instruct-ft.Q5_K_M.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q5_K_M.gguf) | Q5_K_M | 8.64GB |
|
36 |
+
| [Llama-3-13B-Instruct-ft.Q5_1.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q5_1.gguf) | Q5_1 | 9.16GB |
|
37 |
+
| [Llama-3-13B-Instruct-ft.Q6_K.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q6_K.gguf) | Q6_K | 9.98GB |
|
38 |
+
| [Llama-3-13B-Instruct-ft.Q8_0.gguf](https://huggingface.co/RichardErkhov/elinas_-_Llama-3-13B-Instruct-ft-gguf/blob/main/Llama-3-13B-Instruct-ft.Q8_0.gguf) | Q8_0 | 12.92GB |
|
39 |
+
|
40 |
+
|
41 |
+
|
42 |
+
|
43 |
+
Original model description:
|
44 |
+
---
|
45 |
+
base_model:
|
46 |
+
- elinas/Llama-3-13B-Instruct
|
47 |
+
library_name: transformers
|
48 |
+
tags:
|
49 |
+
- mergekit
|
50 |
+
- merge
|
51 |
+
datasets:
|
52 |
+
- Chat-Error/Pure-dove-sharegpt
|
53 |
+
license: llama3
|
54 |
+
---
|
55 |
+
# Llama-3-13B-Instruct-ft
|
56 |
+
|
57 |
+
This is a QLoRA **finetune** of a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit).
|
58 |
+
|
59 |
+
The model is based on my passthrough merge of [Llama-3-13B-Instruct](https://huggingface.co/elinas/Llama-3-13B-Instruct)
|
60 |
+
|
61 |
+
This was primarily an experiment to see how a passthrough merge will respond to further finetuning, though this was done on a small dataset.
|
62 |
+
|
63 |
+
The goal was to make a "mid" sized model like Meta has released in the past and the merge method was inspired by [mlabonne's Llama-3-120B](https://huggingface.co/mlabonne/Meta-Llama-3-120B-Instruct).
|
64 |
+
|
65 |
+
The model was finetuned on **8192 context length** and is likely reliable using RoPE up to 32k.
|
66 |
+
|
67 |
+
It still cannot do math reliably; neither can Llama-3-8B, and in my tests only Llama-3-70B passes basic arithmetic, but it is a better storywriter/RP than Llama-3-8B from some side by side testing I conducted.
|
68 |
+
|
69 |
+
Further finetuning this model or finetuning the [base model](https://huggingface.co/elinas/Llama-3-13B-Instruct) on more samples is encouraged.
|
70 |
+
|
71 |
+
## Datasets
|
72 |
+
|
73 |
+
* [Chat-Error/Pure-dove-sharegpt](https://huggingface.co/datasets/Chat-Error/Pure-dove-sharegpt)
|
74 |
+
|
75 |
+
A small dataset was used to see how it affects performance. Originally I planned to do a larger dataset (196k samples), but wanted to start with a smaller one first to see how much the model improved with some additional finetuning.
|
76 |
+
|
77 |
+
Next steps would be finetuning on a larger dataset if through further testing, performance improvements are noticed.
|
78 |
+
|
79 |
+
## Finetuning details
|
80 |
+
This is a QLoRA model and all modules were targeted.
|
81 |
+
```yaml
|
82 |
+
lora_target_modules:
|
83 |
+
- gate_proj
|
84 |
+
- down_proj
|
85 |
+
- up_proj
|
86 |
+
- q_proj
|
87 |
+
- v_proj
|
88 |
+
- k_proj
|
89 |
+
- o_proj
|
90 |
+
lora_modules_to_save:
|
91 |
+
- embed_tokens
|
92 |
+
- lm_head
|
93 |
+
```
|
94 |
+
|
95 |
+
```yaml
|
96 |
+
The following hyperparameters were used during training:
|
97 |
+
- learning_rate: 1e-05
|
98 |
+
- train_batch_size: 1
|
99 |
+
- eval_batch_size: 1
|
100 |
+
- seed: 42
|
101 |
+
- distributed_type: multi-GPU
|
102 |
+
- num_devices: 3
|
103 |
+
- total_train_batch_size: 3
|
104 |
+
- total_eval_batch_size: 3
|
105 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
106 |
+
- lr_scheduler_type: cosine
|
107 |
+
- lr_scheduler_warmup_steps: 25
|
108 |
+
- num_epochs: 1
|
109 |
+
```
|
110 |
+
|
111 |
+
Optimizer `paged_adamw_8bit` and Deepspeed ZeRO 3 was used at a LR of `1e-5` using the cosine scheduler for 1 epoch on 3x3090s taking 4h 12m 13s total.
|
112 |
+
|
113 |
+
Sample packing and padding was disabled to reduce VRAM consumption significantly at the cost of speed.
|
114 |
+
|
115 |
+
W&B Run Summary
|
116 |
+
```
|
117 |
+
wandb: Run summary:
|
118 |
+
wandb: eval/loss 1.00774
|
119 |
+
wandb: eval/runtime 535.3847
|
120 |
+
wandb: eval/samples_per_second 0.721
|
121 |
+
wandb: eval/steps_per_second 0.241
|
122 |
+
wandb: total_flos 4167452590080.0
|
123 |
+
wandb: train/epoch 1.0
|
124 |
+
wandb: train/global_step 1157
|
125 |
+
wandb: train/grad_norm 4.50846
|
126 |
+
wandb: train/learning_rate 0.0
|
127 |
+
wandb: train/loss 1.4115
|
128 |
+
wandb: train_loss 1.00352
|
129 |
+
wandb: train_runtime 14921.1227
|
130 |
+
wandb: train_samples_per_second 0.233
|
131 |
+
wandb: train_steps_per_second 0.078
|
132 |
+
```
|
133 |
+
|
134 |
+
### Framework versions
|
135 |
+
|
136 |
+
- PEFT 0.10.0
|
137 |
+
- Transformers 4.40.0.dev0
|
138 |
+
- Pytorch 2.3.0+cu121
|
139 |
+
- Datasets 2.15.0
|
140 |
+
- Tokenizers 0.15.0
|
141 |
+
|
142 |
+
## Model Evaluation
|
143 |
+
|
144 |
+
TBD - submitted
|
145 |
+
|
146 |
+
If you have any questions or comments on the model, feel free to open a discussion in the community tab.
|
147 |
+
|
148 |
+
[<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl)
|
149 |
+
|