Possibly made obsolete by: https://huggingface.co/brucethemoose/Yi-34B-200K-DARE-megamerge-v8
Yi 34B 200K DARE Merge v7
A merge of several Yi 34B 200K models using the new DARE Ties method via mergekit. The goal is to create a merge model that excels at 32K+ context performance.
Prompt template: Orca-Vicuna
SYSTEM: {system_message}
USER: {prompt}
ASSISTANT:
It might recognize ChatML, and possibly Alpaca-like formats. Raw prompting as described here is also effective: https://old.reddit.com/r/LocalLLaMA/comments/18zqy4s/the_secret_to_writing_quality_stories_with_llms/
Running
Being a Yi model, try running a lower temperature with 0.02-0.06 MinP, a little repetition penalty, maybe mirostat with a low tau, and no other samplers. Yi tends to run "hot" by default, and it really needs a low temperature + MinP to cull the huge vocabulary.
24GB GPUs can efficiently run Yi-34B-200K models at 45K-90K context with exllamav2, and performant UIs like exui. I go into more detail in this post. 16GB GPUs can still run the high context with aggressive quantization.
To load/train this in full-context backends like transformers, you must change max_position_embeddings
in config.json to a lower value than 200,000, otherwise you will OOM! I do not recommend running high context without context-efficient backends like exllamav2 or unsloth.
Testing Notes
See: https://huggingface.co/brucethemoose/Yi-34B-200K-DARE-merge-v5#testing-notes
A "4k" merge model was created to try and extend the context of SUS Chat and DPO-bagel before adding them to the merge: https://huggingface.co/brucethemoose/SUS-Bagel-200K-DARE-Test
In addition, the weight gradients are biased towards Vicuna-format models in the first few layers to try and "emphasize" the Orca-Vicuna prompt template. How sucessful this is remains to be seen.
Merge Method
This model was merged using the DARE TIES merge method using /home/alpha/Storage/Models/Raw/chargoddard_Yi-34B-200K-Llama as a base.
Models Merged
The following models were included in the merge:
- https://huggingface.co/kyujinpy/PlatYi-34B-200k-Q-FastChat
- https://huggingface.co/jondurbin/bagel-34b-v0.2
- https://huggingface.co/NousResearch/Nous-Capybara-34B
- https://huggingface.co/migtissera/Tess-M-Creative-v1.0
- https://huggingface.co/brucethemoose/SUS-Bagel-200K-DARE-Test
- https://huggingface.co/Mihaiii/Pallas-0.5
- https://huggingface.co/bhenrym14/airoboros-3_1-yi-34b-200k
- https://huggingface.co/adamo1139/Yi-34B-200K-AEZAKMI-v2
- https://huggingface.co/migtissera/Tess-34B-v1.4
- https://huggingface.co/SUSTech/SUS-Chat-34B
- https://huggingface.co/jondurbin/bagel-dpo-34b-v0.2
- https://huggingface.co/chargoddard/Yi-34B-200K-Llama
- https://huggingface.co/chargoddard/Yi-34B-Llama
Configuration
The following YAML configuration was used to produce this model:
models:
- model: /home/alpha/Storage/Models/Raw/chargoddard_Yi-34B-200K-Llama
# No parameters necessary for base model
- model: /home/alpha/Storage/Models/Raw/migtissera_Tess-34B-v1.4
parameters:
weight: [0.23, 0.125, 0.125, 0.125, 0.125, 0.125]
density: 0.59
- model: /home/alpha/Models/Raw/Mihaiii_Pallas-0.5
parameters:
weight: [0.23, 0.125, 0.125, 0.125, 0.125, 0.125]
density: 0.59
- model: /home/alpha//Storage/Models/Raw/bhenrym14_airoboros-3_1-yi-34b-200k
parameters:
weight: [0.02, 0.106, 0.106, 0.106, 0.106, 0.106]
density: 0.59
- model: /home/alpha/Storage/Models/Raw/jondurbin_bagel-34b-v0.2
#Only the SFT in the main merge since the DPO version seems to have no long context ability at all
parameters:
weight: [0.02, 0.100, 0.100, 0.100, 0.100, 0.100]
density: 0.4
- model: /home/alpha/Storage/Models/Raw/kyujinpy_PlatYi-34B-200k-Q-FastChat
parameters:
weight: [0.02, 0.100, 0.100, 0.100, 0.100, 0.100]
density: 0.59
#- model: /home/alpha/Storage/Models/Raw/ehartford_dolphin-2.2-yi-34b-200k
# Dolphin 200K seems to be funky according to multiple leaderboards and perplexity tests?
# parameters:
# weight: 0.15
# density: 0.6
- model: /home/alpha/Models/Raw/adamo1139_Yi-34B-200K-AEZAKMI-v2
parameters:
weight: [0.02, 0.110, 0.110, 0.110, 0.110, 0.110]
density: 0.59
- model: /home/alpha/Storage/Models/Raw/Nous-Capybara-34B
parameters:
weight: [0.22, 0.126, 0.126, 0.126, 0.126, 0.126]
density: 0.59
- model: /home/alpha/Storage/Models/Raw/4kmerge
parameters:
weight: [0.02, 0.108, 0.108, 0.108, 0.108, 0.108]
density: 0.5
- model: /home/alpha/Models/Raw/migtissera_Tess-M-Creative-v1.0
parameters:
weight: [0.22, 0.100, 0.100, 0.100, 0.100, 0.10]
density: 0.59
merge_method: dare_ties
tokenizer_source: union
base_model: /home/alpha/Storage/Models/Raw/chargoddard_Yi-34B-200K-Llama
parameters:
int8_mask: true
dtype: bfloat16
The following config was used for the "4kmerge" model:
models:
- model: /home/alpha/Models/Raw/chargoddard_Yi-34B-Llama
# No parameters necessary for base model
- model: /home/alpha/Storage/Models/Raw/chargoddard_Yi-34B-200K-Llama
parameters:
weight: 0.5
density: 1
- model: /home/alpha/Models/Raw/SUSTech_SUS-Chat-34B
parameters:
weight: 0.2
density: 0.12
- model: /home/alpha/Models/Raw/jondurbin_bagel-dpo-34b-v0.2
parameters:
weight: 0.2
density: 0.15
- model: /home/alpha/Models/Raw/jondurbin_bagel-34b-v0.2
parameters:
weight: 0.1
density: 0.12
merge_method: dare_ties
tokenizer_source: union
base_model: /home/alpha/Models/Raw/chargoddard_Yi-34B-Llama
parameters:
int8_mask: true
dtype: bfloat16
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
Metric | Value |
---|---|
Avg. | 73.12 |
AI2 Reasoning Challenge (25-Shot) | 68.09 |
HellaSwag (10-Shot) | 85.99 |
MMLU (5-Shot) | 77.30 |
TruthfulQA (0-shot) | 58.90 |
Winogrande (5-shot) | 83.11 |
GSM8k (5-shot) | 65.35 |
- Downloads last month
- 3,001
Model tree for brucethemoose/Yi-34B-200K-DARE-merge-v7
Evaluation results
- normalized accuracy on AI2 Reasoning Challenge (25-Shot)test set Open LLM Leaderboard68.090
- normalized accuracy on HellaSwag (10-Shot)validation set Open LLM Leaderboard85.990
- accuracy on MMLU (5-Shot)test set Open LLM Leaderboard77.300
- mc2 on TruthfulQA (0-shot)validation set Open LLM Leaderboard58.900
- accuracy on Winogrande (5-shot)validation set Open LLM Leaderboard83.110
- accuracy on GSM8k (5-shot)test set Open LLM Leaderboard65.350