andreaskoepf
Follow
Highlights
- Pro
Popular repositories
-
faster-rcnn.torch Public
Experimental Torch7 implementation of RCNN for Object Detection with a Region Proposal Network
-
mlp-hidden-visu Public
Visualization of hidden layer activations of small multilayer perceptrons (MLPs)
-
-
-
-
CLIP-Image-Captioning Public
Forked from TheoCoombes/ClipCap
[WIP] Using CLIP and GPT2-XL to generate captions from images.
1,143 contributions in the last year
Contribution activity
April 2023
Created 32 commits in 2 repositories
Created a pull request in LAION-AI/Open-Assistant that received 11 comments
Prepare oasst data export to HuggingFace
Add conversion from trees to 'flat' message table jsonl file.
add oasst-data function to read messages & message tree files read_message_trees, re…
+586
−40
•
11
comments
Opened 15 other pull requests in 1 repository
LAION-AI/Open-Assistant
11
merged
2
open
2
closed
- Add changes for SFT-6 training
- Add gh vars config for rewiew/ranking counts
- Add oa leet10k dataset for SFT training
- Add max_replies parameter for trainer_rm
- Fix export_model script
- Fix 'ConfigError: Attempted to change value of key "val_max_length" ...'
- Add option to export trlx checkpoints
- Model training dockerfile
- Add HFSummaryPairs class & fix AnthropicRLHF parsing
- Improve scores of small 1.4B reward model..
- Use fixed RNG seed value for all DeepSpeed workers
- Code state of wandb run bqiatai0 (2023-03-25_oasst_cyrillic_alpaca_reference_good)
- ignore, see #2312 for correct version
- Revert unrelated changes in instructor rank_datasets
- Fix loading of Nebulous/gpt4all_pruned dataset
Reviewed 49 pull requests in 1 repository
LAION-AI/Open-Assistant
25 pull requests
- add draft OpenAssistant OASST1 data notebook
- Added yk to website owners and fixed a path bug
- Fixed worker requirements wrt transformers
- Update oasst-data docs to reflect code changes
- Added "current announcement" to web (minimal implementation)
- Made "running locally" more precise
- Reduced k50 temperature to 0.9
- Playbook string env fix
- Add release video as blog post
- Gunicorn worker image for inference
- Remove safety deployment dependency in release playbook
- Added enable inference safety to playbook
- remove references from dolly
- Added playbook variables for google auth
- add databricks dolly dataset
- Added llama 30b sft-5 to model configs
- Add oa leet10k dataset for SFT training
- Add Databricks Dolly 15k converted to OA data format
- Add sedthh to data code owners
- add Logic Inference Dataset
- remove openai references from datasets
- fix list index out of range error for soda dataset
- Fix for Webgpt NaN loss
- Improvements to the HF worker container
- Added docker image for standalone worker
- Some pull request reviews not shown.
Created an issue in LAION-AI/Open-Assistant that received 3 comments
Markdown homogenization of training data
Some of our datasets are markdown formatted and others in plain-text. Datasets using strictly markdown (e.g. after conversion from html with a tool…
3
comments
Opened 8 other issues in 1 repository
LAION-AI/Open-Assistant
7
open
1
closed
- Add warning message near chat window about model hallucinations
- Fill <|system|> prompt during supervised fine-tuning with lang, len & text-labels
- Add reward model scoring during SFT evaluation
- Evaluate Microsoft DeepSpeed-Chat Training System
- Consolidate two different RL trainer implementations into OA main repo
- Integrate reward/instructor implementation into trainer_rm
- Update trainer_rm training code to correctly use different rank training sets
- SFT-Training with DeepSpeed 0.8.3 produces significantly worse results than with 0.7.7