Commit Graph

112 Commits

Author SHA1 Message Date
Googler 066f229e27 fix(rlhf): Supporting adapter only output for reward model training
PiperOrigin-RevId: 608740017
2024-02-20 14:27:59 -08:00
Googler fc183f3acb chore(components): Rename several `_implementation.llm` components
PiperOrigin-RevId: 607487816
2024-02-15 16:18:27 -08:00
Michael Hu 449c304686 fix(components): Use PipelineJob location in AutoSxS components, add init file
PiperOrigin-RevId: 607055407
2024-02-14 11:40:06 -08:00
Googler 0b75afdd8a chore(components): Update AutoSxS and RLHF image tags
PiperOrigin-RevId: 606330905
2024-02-12 12:30:46 -08:00
Michael Hu 1280753eb4 chore(components): Use new module for looking up ReFINED and AutoSxS image tags
PiperOrigin-RevId: 605457014
2024-02-08 16:14:35 -08:00
Googler d4c3f35797 feat(components): Add RLAIF pipeline to preview
PiperOrigin-RevId: 605396378
2024-02-08 12:36:38 -08:00
Googler 87db18e3a1 No public description
PiperOrigin-RevId: 605012378
2024-02-07 09:41:59 -08:00
Michael Hu b6247fb8e4 chore(components): Update component naming in AutoSxS implementation
PiperOrigin-RevId: 604678269
2024-02-06 09:49:04 -08:00
Michael Hu 14193def65 chore(components): Create module containing AutoSxS and RLHF image tag
PiperOrigin-RevId: 603765313
2024-02-02 12:58:35 -08:00
Michael Hu 8c7b5b2bf5 feat(components): Use a single inference component for AutoSxS
PiperOrigin-RevId: 601877680
2024-01-26 15:06:46 -08:00
Googler b9e08ded48 fix(components): Only run `preview.llm.bulk_inference` after tuning third-party models with RLHF
PiperOrigin-RevId: 601226133
2024-01-24 13:59:24 -08:00
Googler ddb2f9a8b6 fix(components): Update base image for KFP lightweight component for VPC SC compliance
PiperOrigin-RevId: 600622671
2024-01-22 17:55:51 -08:00
Googler 60b66dca0f docs(components): Update AutoSxS pipeline to use "question_answering" as task name instead of "question_answer", where "question_answer" is still supported, but deprecated
chore(components): Update RLHF and AutoSxS image tags

PiperOrigin-RevId: 600532640
2024-01-22 12:17:59 -08:00
Googler 4bb3423889 feat(components): Support scheduling and labels in utils.build_payload
PiperOrigin-RevId: 599904659
2024-01-19 12:20:10 -08:00
Jason Dai 80c9b04bd6 fix(components): Update base image for KFP lightweight component for VPC SC compliance
PiperOrigin-RevId: 598888543
2024-01-16 10:46:58 -08:00
Googler 4e1491afd6 feat(components): Implement new output format of inference component
PiperOrigin-RevId: 597621035
2024-01-11 12:12:18 -08:00
Googler d73c6db3de feat(components): Implement the feature store grounding pipeline
PiperOrigin-RevId: 597032159
2024-01-09 13:42:23 -08:00
Googler a66c5990e4 feat(components): Output errors as a separate table from Arbiter
PiperOrigin-RevId: 592769441
2023-12-21 00:03:12 -08:00
Michael Hu 14de087e74 No public description
PiperOrigin-RevId: 592702825
2023-12-20 17:53:15 -08:00
Googler 44f9992d0c feat(components): change output format to allow possible post eval
PiperOrigin-RevId: 592257929
2023-12-19 09:25:57 -08:00
Googler 8eddc27bc9 chore(components): Fix argument description
PiperOrigin-RevId: 592033936
2023-12-18 15:56:52 -08:00
Googler 075d58f89f fix(components): Resolve unique model display name on each `preview.llm.rlhf_pipeline` run instead of reusing cached result
PiperOrigin-RevId: 591365087
2023-12-15 14:48:17 -08:00
Googler f51a930120 fix(components): Use `large_model_reference` as `model_reference_name` when uploading models from `preview.llm.rlhf_pipeline` instead of hardcoding value as `text-bison@001`
PiperOrigin-RevId: 591346782
2023-12-15 13:36:40 -08:00
Jason Dai 8c59816bf2 feat(components): Release new LLM Eval image version 0.5
PiperOrigin-RevId: 590702775
2023-12-13 13:53:17 -08:00
Googler 3d62d26727 feat(component): Migrate AutoSxS pipeline to preview and move related files to _implementation/llm directory to help Model Eval team use side by side metrics as part of their pipeline
PiperOrigin-RevId: 590677273
2023-12-13 12:20:20 -08:00
Googler efeed83406 feat(components): Enable text generation pipeline to generate row based metrics
docs: update description of ModelEvaluation's metadata field accordingly

PiperOrigin-RevId: 590486026
2023-12-13 00:43:57 -08:00
Googler 227eab1c68 fix(components): Use `llama-2-7b` for the base reward model when tuning `llama-2-13` with the `preview.llm.rlhf_pipeline`
PiperOrigin-RevId: 589490121
2023-12-09 18:23:01 -08:00
Googler 685634d4a3 feat(components): Add `num_microbatches` to `_implementation.llm` training components
PiperOrigin-RevId: 589267101
2023-12-08 16:07:09 -08:00
Googler 9007fb0007 feat(components): Bump image tag used by `preview.llm` pipelines
PiperOrigin-RevId: 589253163
2023-12-08 15:10:38 -08:00
Googler 708b8bd623 fix(components): Append `tune-type` label when uploading models tuned by `preview.llm.rlhf_pipeline`
PiperOrigin-RevId: 588172611
2023-12-05 13:30:22 -08:00
Googler 1f698349f1 feat(components): Enable endpoint_batch_predict component to take in publisher model by either model_address + model_id or just model_id
PiperOrigin-RevId: 586812440
2023-11-30 15:56:05 -08:00
Googler 397b1c97be feat(components): Edit embedding pipeline to use generic batch predict component to support multilingual embeddings model
PiperOrigin-RevId: 586329396
2023-11-29 07:01:49 -08:00
Googler 1e089e6f6d feat(components): [third party model inference] Initialize component
PiperOrigin-RevId: 585766918
2023-11-27 14:56:55 -08:00
Jason Dai a281dd9935 chore(components): release new LLM Eval container image `v0.4`
PiperOrigin-RevId: 584398807
2023-11-21 12:27:54 -08:00
Googler 69446b7817 chore(components): [endpoint_batch_predict] fix pipeline description
PiperOrigin-RevId: 583079495
2023-11-16 09:56:31 -08:00
Googler c23b720f10 feat(components): Group `preview.llm.rlhf_pipeline` components for more readability
PiperOrigin-RevId: 582733817
2023-11-15 10:56:39 -08:00
Jason Dai 6cfad2b348 feat(components): Create new eval dataset preprocessor for formatting eval dataset
PiperOrigin-RevId: 582731483
2023-11-15 10:47:02 -08:00
Googler c038ef0a5b chore(components): [endpoint_batch_predict] update image uri to vertex-evaluation for fishfooding
PiperOrigin-RevId: 582725363
2023-11-15 10:29:05 -08:00
Googler bcd59220f4 feat(components): Group `preview.llm.rlhf_pipeline` components for more readability
PiperOrigin-RevId: 582394326
2023-11-14 11:52:02 -08:00
Googler a927984394 feat(components): Group `preview.llm.rlhf_pipeline` components for more readability
PiperOrigin-RevId: 581354060
2023-11-10 13:32:31 -08:00
Googler 71e5a938ef feat(components):[text2sql] Integration with first party LLM model inference pipeline
PiperOrigin-RevId: 581300725
2023-11-10 10:27:54 -08:00
Googler a54ac75c8c feat(components): [endpoint_batch_predict] add retry function for retryable error
PiperOrigin-RevId: 580980127
2023-11-09 11:52:58 -08:00
Googler 4a5cbbfb8d feat(components): Update image tag used by RLHF components
PiperOrigin-RevId: 580698126
2023-11-08 16:27:14 -08:00
Jason Dai f6aad5e4e0 feat(components): add LLM Eval pipeline parameter for customizing eval dataset reference ground truth field
PiperOrigin-RevId: 580233829
2023-11-07 10:54:21 -08:00
Googler a8dd3117d5 No public description
PiperOrigin-RevId: 579961750
2023-11-06 14:40:08 -08:00
Googler 99b67ea426 chore(components): [endpoint_batch_predict] update input and output interface so that it's consistent with existing batch predict service
PiperOrigin-RevId: 579255083
2023-11-03 12:08:49 -07:00
Googler d9c3b86adc chore(components): [endpoint_batch_predict] add qms_override parameter data type conversion
PiperOrigin-RevId: 578698352
2023-11-01 18:08:16 -07:00
Googler ed9a5abe3a chore(components): [endpoint_batch_predict] update pipeline name
PiperOrigin-RevId: 578361190
2023-10-31 18:26:05 -07:00
Googler 2910d0bb52 feat(components):[text2sql] Generate SQL queries by model batch prediction
PiperOrigin-RevId: 578340803
2023-10-31 16:46:27 -07:00
Googler 9b7a5dc532 chore(components): [endpoint_batch_predict] update parameters' default value
PiperOrigin-RevId: 578266253
2023-10-31 12:28:06 -07:00