This repository has been archived by the owner on Nov 3, 2023. It is now read-only.
New Releases
- K2R: Reason first, then respond: Modular Generation for Knowledge-infused Dialogue (#4828)
- ROSCOE: A Suite of Metrics for Scoring Step-by-Step Reasoning (#4839, #4856, #4860, #4862)
- The CRINGE Loss: Learning what language not to model (#4871, #4876, #5036)
- Multi-party LIGHT: Multi-Party Chat: Conversational Agents in Group Settings with Humans and Models (aka MultiLIGHT) (#5043, #5041, #5045)
New Features
- New releases coming with a public Docker image (#4868, #4874)
- flan-t5 models in ParlAI with fp16 (#4875)
- Factual Nucleus Sampling (#4890)
- [FSDP] Zero 3 Optimization Support (#4903, #4911)
- [BB3] New Module; New Customization (#4944, #4946)
- Workers' qualifications for the model chat (#4956)
- ClearML Logger (#4896)
- Async multi-party model chat (#4993)
- jsonfile retains original fields (#5033)
Metrics
Bug fixes
GRM: (#4796)- Module Level Tasks (#4798)
- Display data with utf-16 (#4864) and emojis (#4833, #4934 (revert #4666))
- Reasoning task (#4883)
- TGA Compatibility with Pytorch>=1.13 (#4887)
- T5 Init Bug (#4897)
- Delay Loading NGram-Blocking on GPU (#4886)
- Fix starspace (#5003)
- Datatype option for Wizard of Internet task (#4962)
- T5 and GPT2 fixes (#5016, #5032)
Datasets & Teachers
- Names Bias (#4836)
Developers & Documentation
- Minor fixes and typos (#4841, #4873, #4879, #4892, #4901, #4902, #4912, #4947, #4949, #4954, #4959, #4975, #4987, #5026, #5044)
- Docker usage tutorial (#4874)
- Installation Instructions #5015