Comments (8)
Line 1252 in 97b9f2d
(env) arno@rippa:/nfs4/llm/h2ogpt(main)$ NCCL_P2P_LEVEL=LOC WORLD_SIZE=2 CUDA_VISIBLE_DEVICES="0,1" torchrun --nproc_per_node=2 finetune.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6.9b --data_path=h2oai/openassistant_oasst1_h2ogpt_only --micro_batch_size=2 --batch_size=32 --cutoff_len=2048 --run_id=100 --num_epochs=1 &> run.100.text
https://slack-files.com/T0329MHH6-F054U3491DF-a909f3bb3b lora weights and logs
from h2ogpt.
Comparing validation score on only 1k ShareGPT before/after personality refresh, so just "hammering" the personality updates, and nothing else.
before:
CUDA_VISIBLE_DEVICES=1 python generate.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6.9b --prompt_type='human_bot' --chat=False --stream_output=False --gradio=False --eval_sharegpt_prompts_only=1000 --eval_sharegpt_as_output=False --num_beams=1 &> h2ogpt-oig-oasst1-512-6.9b.eval.log
h2ogpt-oig-oasst1-512-6.9b.eval.log
after:
CUDA_VISIBLE_DEVICES=0 python generate.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6.9b --lora_weights=h2ogpt-oig-oasst1-512-6.9b.h2oaiopenassistant_oasst1_h2ogpt_only.1_epochs.97b9f2df6851e8b1c2f220a1028adad9029140bf.100 --prompt_type='human_bot' --chat=False --stream_output=False --gradio=False --eval_sharegpt_prompts_only=1000 --eval_sharegpt_as_output=False --num_beams=1 &> h2ogpt-oig-oasst1-512-6.9b.h2oaiopenassistant_oasst1_h2ogpt_only.1_epochs.97b9f2df6851e8b1c2f220a1028adad9029140bf.100.eval.log
h2ogpt-oig-oasst1-512-6.9b.h2oaiopenassistant_oasst1_h2ogpt.2_epochs.e35e2e06e0af2f7dceac2e16e3646c90ccce4ec0.1.eval.log
from h2ogpt.
give h2oai/h2ogpt-oasst1-512-6.9b personality
So obviously hurts too much to only hammer in the personality and nothing else, need to keep variety of fine-tuning data, so easiest for smaller models is to train a couple epochs on the full personalized dataset instead
https://huggingface.co/datasets/h2oai/openassistant_oasst1_h2ogpt
For larger models, can try to do some mixin mode, with small personality data + large mixin, but not all 80k rows. Note: didn't do that, just added 2 epochs on top of existing *512*
models.
torchrun --nproc_per_node=2 finetune.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6.9b --data_path=h2oai/openassistant_oasst1_h2ogpt --prompt_type=plain --run_id=1 --micro_batch_size=8 --batch_size=512 --cutoff_len=512 --num_epochs=2 &> log.1.txt
https://slack-files.com/T0329MHH6-F054GHUJJFM-eebaba5e98 lora weights and logs
https://huggingface.co/h2oai/h2ogpt-oig-oasst1-512-6.9b model card created by fcaae7e
from h2ogpt.
CUDA_VISIBLE_DEVICES=0 python generate.py --base_model=h2oai/h2ogpt-oig-oasst1-512-6.9b --lora_weights=h2ogpt-oig-oasst1-512-6.9b.h2oaiopenassistant_oasst1_h2ogpt.2_epochs.e35e2e06e0af2f7dceac2e16e3646c90ccce4ec0.1 --prompt_type='human_bot' --chat=False --stream_output=False --gradio=False --eval_sharegpt_prompts_only=1000 --eval_sharegpt_as_output=False --num_beams=1 &> h2ogpt-oig-oasst1-512-6.9b.h2oaiopenassistant_oasst1_h2ogpt.2_epochs.e35e2e06e0af2f7dceac2e16e3646c90ccce4ec0.1.eval.log
from h2ogpt.
give h2oai/h2ogpt-oasst1-512-12b personality
continue #22 (comment)
torchrun --nproc_per_node=2 finetune.py --base_model=h2oai/h2ogpt-oasst1-512-12b --data_path=h2oai/openassistant_oasst1_h2ogpt --prompt_type=plain --run_id=2 --micro_batch_size=4 --batch_size=128 --cutoff_len=512 --num_epochs=2 &> log.2.txt
from h2ogpt.
Comparing validation score on 1k ShareGPT before/after personality refresh
Before:
CUDA_VISIBLE_DEVICES=0,1 python generate.py --base_model=h2oai/h2ogpt-oasst1-512-12b --prompt_type='human_bot' --infer_devices=False --chat=False --stream_output=False --gradio=False --eval_sharegpt_prompts_only=1000 --eval_sharegpt_as_output=False --num_beams=1 &> h2ogpt-oasst1-512-12b.eval.log
h2ogpt-oasst1-512-12b.eval.log
After:
CUDA_VISIBLE_DEVICES=0,1 python generate.py --base_model=h2oai/h2ogpt-oasst1-512-12b --lora_weights=h2ogpt-oasst1-512-12b.h2oaiopenassistant_oasst1_h2ogpt.2_epochs.fcaae7ef70600de8c97c9b38cb3f0075467cdad1.2 --prompt_type='human_bot' --infer_devices=False --chat=False --stream_output=False --gradio=False --eval_sharegpt_prompts_only=1000 --eval_sharegpt_as_output=False --num_beams=1 &> h2ogpt-oasst1-512-12b.h2oaiopenassistant_oasst1_h2ogpt.2_epochs.fcaae7ef70600de8c97c9b38cb3f0075467cdad1.2.eval.log
h2ogpt-oasst1-512-12b.h2oaiopenassistant_oasst1_h2ogpt.2_epochs.fcaae7ef70600de8c97c9b38cb3f0075467cdad1.2.eval.log
from h2ogpt.
give h2oai/h2ogpt-oasst1-512-20b personality
torchrun --nproc_per_node=8 finetune.py --base_model=h2oai/h2ogpt-oasst1-512-20b --data_path=h2oai/openassistant_oasst1_h2ogpt --prompt_type=plain --run_id=3 --micro_batch_size=4 --batch_size=128 --cutoff_len=512 --num_epochs=2 &> log.3.txt
https://slack-files.com/T0329MHH6-F054HBLG2LB-7a39e222f8 lora weights and logs
h2oai/h2ogpt-oasst1-512-20b
e0202c5
from h2ogpt.
from h2ogpt.
Related Issues (20)
- Shared collection problem with multiple users HOT 6
- "configure_parser" issue at install HOT 3
- Add JSON et al. for output, e.g. guided_json if vllm, openai way, mistralai for models supported, etc.
- Does this tool guarantee that all data is private? HOT 2
- Can you add langsmith/wandb for tracing and ragas for evaluation metrics? HOT 1
- Mac OS manual Installation runs error HOT 2
- Mac OS automatic installation runs errors HOT 2
- about Add Doc to Chat HOT 1
- One Click Installers for MacOS not working on MacMini M2
- Attention sink error with h2oai/h2ogpt-4096-llama2-13b-chat HOT 1
- ImportError: libcudnn.so.8: cannot open shared object file: No such file or directory HOT 6
- Size of Tensor A must match size of Tensor B HOT 6
- auth related feature HOT 9
- Loading a Large model on Multiples GPU system HOT 12
- Permissions in VectorDB HOT 6
- Support for AWS Bedrock HOT 1
- vLLM GROQ issue HOT 1
- Mac OS auto installer doesn't work after manual uninstallation
- RuntimeError: An error occurred while downloading using `hf_transfer`. HOT 1
- python dependency module version tweaks HOT 1
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from h2ogpt.