Hugging Face's logo

PJMixers-Dev
/
LLaMa-3.2-Instruct-JankMix-v0.2-SFT-HailMary-v0.1-KTO-3B-GGUF

GGUF
English
Eval Results
Inference Endpoints
conversational
Model card Files Community
xzuyn commited on
Commit
7ba4754
•
1 Parent(s): cb0b286

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +130 -0
README.md ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: llama3.2
3
+ language:
4
+ - en
5
+ datasets:
6
+ - PJMixers-Dev/HailMary-v0.1-KTO
7
+ base_model:
8
+ - PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-3B
9
+ model-index:
10
+ - name: PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-HailMary-v0.1-KTO-3B
11
+ results:
12
+ - task:
13
+ type: text-generation
14
+ name: Text Generation
15
+ dataset:
16
+ name: IFEval (0-Shot)
17
+ type: HuggingFaceH4/ifeval
18
+ args:
19
+ num_few_shot: 0
20
+ metrics:
21
+ - type: inst_level_strict_acc and prompt_level_strict_acc
22
+ value: 65.04
23
+ name: strict accuracy
24
+ source:
25
+ url: https://huggingface.co/datasets/open-llm-leaderboard/PJMixers-Dev__LLaMa-3.2-Instruct-JankMix-v0.2-SFT-HailMary-v0.1-KTO-3B-details
26
+ name: Open LLM Leaderboard
27
+ - task:
28
+ type: text-generation
29
+ name: Text Generation
30
+ dataset:
31
+ name: BBH (3-Shot)
32
+ type: BBH
33
+ args:
34
+ num_few_shot: 3
35
+ metrics:
36
+ - type: acc_norm
37
+ value: 22.29
38
+ name: normalized accuracy
39
+ source:
40
+ url: https://huggingface.co/datasets/open-llm-leaderboard/PJMixers-Dev__LLaMa-3.2-Instruct-JankMix-v0.2-SFT-HailMary-v0.1-KTO-3B-details
41
+ name: Open LLM Leaderboard
42
+ - task:
43
+ type: text-generation
44
+ name: Text Generation
45
+ dataset:
46
+ name: MATH Lvl 5 (4-Shot)
47
+ type: hendrycks/competition_math
48
+ args:
49
+ num_few_shot: 4
50
+ metrics:
51
+ - type: exact_match
52
+ value: 11.78
53
+ name: exact match
54
+ source:
55
+ url: https://huggingface.co/datasets/open-llm-leaderboard/PJMixers-Dev__LLaMa-3.2-Instruct-JankMix-v0.2-SFT-HailMary-v0.1-KTO-3B-details
56
+ name: Open LLM Leaderboard
57
+ - task:
58
+ type: text-generation
59
+ name: Text Generation
60
+ dataset:
61
+ name: GPQA (0-shot)
62
+ type: Idavidrein/gpqa
63
+ args:
64
+ num_few_shot: 0
65
+ metrics:
66
+ - type: acc_norm
67
+ value: 2.91
68
+ name: acc_norm
69
+ source:
70
+ url: https://huggingface.co/datasets/open-llm-leaderboard/PJMixers-Dev__LLaMa-3.2-Instruct-JankMix-v0.2-SFT-HailMary-v0.1-KTO-3B-details
71
+ name: Open LLM Leaderboard
72
+ - task:
73
+ type: text-generation
74
+ name: Text Generation
75
+ dataset:
76
+ name: MuSR (0-shot)
77
+ type: TAUR-Lab/MuSR
78
+ args:
79
+ num_few_shot: 0
80
+ metrics:
81
+ - type: acc_norm
82
+ value: 4.69
83
+ name: acc_norm
84
+ source:
85
+ url: https://huggingface.co/datasets/open-llm-leaderboard/PJMixers-Dev__LLaMa-3.2-Instruct-JankMix-v0.2-SFT-HailMary-v0.1-KTO-3B-details
86
+ name: Open LLM Leaderboard
87
+ - task:
88
+ type: text-generation
89
+ name: Text Generation
90
+ dataset:
91
+ name: MMLU-PRO (5-shot)
92
+ type: TIGER-Lab/MMLU-Pro
93
+ config: main
94
+ split: test
95
+ args:
96
+ num_few_shot: 5
97
+ metrics:
98
+ - type: acc
99
+ value: 23.42
100
+ name: accuracy
101
+ source:
102
+ url: https://huggingface.co/datasets/open-llm-leaderboard/PJMixers-Dev__LLaMa-3.2-Instruct-JankMix-v0.2-SFT-HailMary-v0.1-KTO-3B-details
103
+ name: Open LLM Leaderboard
104
+ ---
105
+ [PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-3B](https://huggingface.co/PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-3B) was further trained using KTO (with `apo_zero_unpaired` loss type) using a mix of instruct, RP, and storygen datasets. I created rejected samples by using the SFT with bad settings (including logit bias) for every model turn.
106
+
107
+ The model was only trained at `max_length=6144`, and is nowhere near a full epoch as it eventually crashed. So think of this like a test of a test.
108
+
109
+ # W&B Training Logs
110
+
111
+ ![train/rewards/chosen/rejected](https://huggingface.co/PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-HailMary-v0.1-KTO-3B/resolve/main/images/train_rewards_chosen_rejected.png)
112
+ ![train/rewards/margins](https://huggingface.co/PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-HailMary-v0.1-KTO-3B/resolve/main/images/train_rewards_margins.png)
113
+ ![train/logits/chosen/rejected](https://huggingface.co/PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-HailMary-v0.1-KTO-3B/resolve/main/images/train_logits_chosen_rejected.png)
114
+ ![train/logps/chosen/rejected](https://huggingface.co/PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-HailMary-v0.1-KTO-3B/resolve/main/images/train_logps_chosen_rejected.png)
115
+ ![train/loss](https://huggingface.co/PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-HailMary-v0.1-KTO-3B/resolve/main/images/train_loss.png)
116
+ ![train/grad_norm](https://huggingface.co/PJMixers-Dev/LLaMa-3.2-Instruct-JankMix-v0.2-SFT-HailMary-v0.1-KTO-3B/resolve/main/images/train_grad_norm.png)
117
+
118
+ # [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard)
119
+
120
+ Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/PJMixers-Dev__LLaMa-3.2-Instruct-JankMix-v0.2-SFT-3B-details)
121
+
122
+ | Metric |Value|
123
+ |-------------------|----:|
124
+ |Avg. |21.69|
125
+ |IFEval (0-Shot) |65.04|
126
+ |BBH (3-Shot) |22.29|
127
+ |MATH Lvl 5 (4-Shot)|11.78|
128
+ |GPQA (0-shot) | 2.91|
129
+ |MuSR (0-shot) | 4.69|
130
+ |MMLU-PRO (5-shot) |23.42|