File size: 1,624 Bytes
f9bc2c5
 
 
 
 
 
 
 
 
 
fc8053b
f9bc2c5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: other
license_name: llama-3
license_link: https://llama.meta.com/llama3/license/
tags:
- llama-3
- llama
- '3'
- 5B
---
This is just an experiment similar to that done on [chargoddard/llama3-42b-v0](https://huggingface.co/chargoddard/llama3-42b-v0). The post-pruning was fine-tuned or "healed" with QLoRA using the code DPO dataset [AlekseyKorshuk/evol-codealpaca-v1-dpo](https://huggingface.co/datasets/AlekseyKorshuk/evol-codealpaca-v1-dpo). Due to limitations, this was only trained on 3150/4935 (~64%) steps of the data. I had to restart the training about halfway through, so the logs are split in two. I am still unsure if the tokenizer is correct.

Loss: ~1.2
<img src="https://i.imgur.com/AnuMlv7.png">

<img src="https://i.imgur.com/kHXnKCU.png">

<img src="https://i.imgur.com/aHKVgqT.png">

<img src="https://i.imgur.com/KTLYnjl.png">

mergekit.yaml
```
slices:
  - sources:
      - model: ./Meta-Llama-3-8B-Instruct/
        layer_range: [0,15]
  - sources:
      - model: ./Meta-Llama-3-8B-Instruct/
        layer_range: [29,32]
            
merge_method: passthrough
dtype: bfloat16
```

ORPOConfig
```
    learning_rate=5e-5,
    lr_scheduler_type="cosine",
    max_length=1024,
    max_prompt_length=512,
    overwrite_output_dir=False,
    beta=0.1,
    per_device_train_batch_size=2,
    per_device_eval_batch_size=2,
    gradient_accumulation_steps=4,
    optim="paged_adamw_8bit",
    num_train_epochs=1,
    evaluation_strategy="steps",
    eval_steps=0.02,
    logging_steps=1,
    warmup_steps=50,
    report_to="wandb",
    output_dir=out_dir_folder,
    fp16=True,
    save_steps=50
```