Code

RL math evaluation#109

Code

Merged

AlexPiche merged 58 commits intomainServiceNow/TapeAgents:mainfrom rl_for_mathServiceNow/TapeAgents:rl_for_mathCopy head branch name to clipboardDec 9, 2024

+1,594-291

Commits

Commits on Nov 20, 2024

• typo

  AlexPiche

  committed
  a8bdf09

  Copy full SHA for a8bdf09

Commits on Nov 21, 2024

• simplify code

  AlexPiche

  committed
  e5477be

  Copy full SHA for e5477be

• clean up

  AlexPiche

  committed
  2c0e603

  Copy full SHA for 2c0e603

• clean up orchestrate

  AlexPiche

  committed
  dd638f1

  Copy full SHA for dd638f1

Commits on Nov 22, 2024

• use deepseek math evals

  AlexPiche

  committed
  5c5c44e

  Copy full SHA for 5c5c44e

• math evals

  AlexPiche

  committed
  6af2a30

  Copy full SHA for 6af2a30

• reverse changes to gsm8k example

  AlexPiche

  committed
  9c8e6ff

  Copy full SHA for 9c8e6ff

• Merge remote-tracking branch 'origin/main' into rl_for_math

  AlexPiche

  committed
  6e6a722

  Copy full SHA for 6e6a722

• clean up

  AlexPiche

  committed
  88e68c7

  Copy full SHA for 88e68c7

• refactor deepseek evals

  AlexPiche

  committed
  2162c3f

  Copy full SHA for 2162c3f

• mit license

  AlexPiche

  committed
  50ba5b8

  Copy full SHA for 50ba5b8

• license

  AlexPiche

  committed
  fb9cc92

  Copy full SHA for fb9cc92

• license

  AlexPiche

  committed
  90a819d

  Copy full SHA for 90a819d

• clean up

  AlexPiche

  committed
  7b5eb46

  Copy full SHA for 7b5eb46

• handle empty list

  AlexPiche

  committed
  3a10fa5

  Copy full SHA for 3a10fa5

• handle empty list

  AlexPiche

  committed
  d9d9ca0

  Copy full SHA for d9d9ca0

• clean up node

  AlexPiche

  committed
  14e082b

  Copy full SHA for 14e082b

• fix answers separated by commas

  AlexPiche

  committed
  0f72fda

  Copy full SHA for 0f72fda

Commits on Nov 23, 2024

• handle comma separated answers

  AlexPiche

  committed
  4e8caa7

  Copy full SHA for 4e8caa7

Commits on Nov 25, 2024

• clip advantage

  AlexPiche

  committed
  ad08046

  Copy full SHA for ad08046

• do not raise for mismatch tokens

  AlexPiche

  committed
  3db15e1

  Copy full SHA for 3db15e1

• max_seq_len inherits from max tokens

  AlexPiche

  committed
  c19d194

  Copy full SHA for c19d194

Commits on Nov 26, 2024

• llm call output tokens were off by 4

  AlexPiche

  committed
  9e0fe98

  Copy full SHA for 9e0fe98

Commits on Nov 27, 2024

• better logging

  AlexPiche

  committed
  ed4b187

  Copy full SHA for ed4b187

• path for deepseek model

  AlexPiche

  committed
  9d4ce1e

  Copy full SHA for 9d4ce1e

• fix adv clip

  AlexPiche

  committed
  f456c95

  Copy full SHA for f456c95

Commits on Nov 28, 2024

• non accelerate training

  AlexPiche

  committed
  d961fb5

  Copy full SHA for d961fb5

• revert to adamw optim

  AlexPiche

  committed
  ec672eb

  Copy full SHA for ec672eb

• fix reward norm

  AlexPiche

  committed
  4df988b

  Copy full SHA for 4df988b

• fix reward norm

  AlexPiche

  committed
  914820e

  Copy full SHA for 914820e

Commits on Nov 30, 2024

• fix adamw

  AlexPiche

  committed
  97313bc

  Copy full SHA for 97313bc

Commits on Dec 2, 2024

• test on the full set

  AlexPiche

  committed
  b7f669e

  Copy full SHA for b7f669e

• Merge branch 'main' into rl_for_math

  AlexPiche

  authored
  a4e6204

  Copy full SHA for a4e6204

• write adamw comment

  AlexPiche

  committed
  306d01d

  Copy full SHA for 306d01d

• Merge branch 'rl_for_math' of github.com:ServiceNow/TapeAgents into rl_for_math

  AlexPiche

  committed
  84360ee

  Copy full SHA for 84360ee

• import multiprocessing

  AlexPiche

  committed
  1811bb4

  Copy full SHA for 1811bb4

• run finetune loop

  AlexPiche

  committed
  40c32ed

  Copy full SHA for 40c32ed

• fix litellm import

  AlexPiche

  committed
  d087d4c

  Copy full SHA for d087d4c

• revert change

  AlexPiche

  committed
  3d8ff86

  Copy full SHA for 3d8ff86

Commits on Dec 3, 2024

• Merge remote-tracking branch 'origin/save_logprobs_to_sqlite' into rl_for_math

  AlexPiche

  committed
  7025e7a

  Copy full SHA for 7025e7a

• fix mathtape symbol

  AlexPiche

  committed
  cfac027

  Copy full SHA for cfac027

Commits on Dec 4, 2024

• Merge remote-tracking branch 'origin/save_logprobs_to_sqlite' into rl_for_math

  AlexPiche

  committed
  87f9777

  Copy full SHA for 87f9777

• cardMerge remote-tracking branch 'origin/save_logprobs_to_sqlite' into rl_for_math

  AlexPiche

  committed
  d8206d1

  Copy full SHA for d8206d1

• Merge branch 'save_logprobs_to_sqlite' into rl_for_math

  AlexPiche

  authored
  3c0d59f

  Copy full SHA for 3c0d59f

• merge conflicts

  AlexPiche

  committed
  2f7e935

  Copy full SHA for 2f7e935

• rm leading white space

  AlexPiche

  committed
  3794553

  Copy full SHA for 3794553

Commits on Dec 5, 2024

• merge conflicts

  AlexPiche

  committed
  f2881f6

  Copy full SHA for f2881f6

• revert lstrip from chat_template

  AlexPiche

  committed
  33b6d42

  Copy full SHA for 33b6d42

• clean up

  AlexPiche

  committed
  49ec5c9

  Copy full SHA for 49ec5c9

• rm adv clipping

  AlexPiche

  committed
  45f7f94

  Copy full SHA for 45f7f94

• fix content

  AlexPiche

  committed
  dced6fb

  Copy full SHA for dced6fb

Commits on Dec 6, 2024

• rm non accelerate support

  AlexPiche

  committed
  ed3a450

  Copy full SHA for ed3a450

Commits on Dec 7, 2024

• Merge remote-tracking branch 'origin/save_logprobs_to_sqlite' into rl_for_math

  AlexPiche

  committed
  c44b0cc

  Copy full SHA for c44b0cc

• fix conflicts

  AlexPiche

  committed
  06ef599

  Copy full SHA for 06ef599

Commits on Dec 9, 2024

• merge main

  AlexPiche

  committed
  be6f6df

  Copy full SHA for be6f6df

• rm useless logging

  AlexPiche

  committed
  1656d31

  Copy full SHA for 1656d31

• formatting

  AlexPiche

  committed
  f592573

  Copy full SHA for f592573

• fix rl test

  AlexPiche

  committed
  b6bf0fb

  Copy full SHA for b6bf0fb