Commit graph

10 commits

Author SHA1 Message Date
Zafir Stojanovski
a969d8ef05
feat(curriculum): Knights and Knaves configs (#488)
* configs

* reduce complexity of curriculum

* update lower bound

* add failure threshold

* update last_k

* update thresholds for success and failure

* update curriculum file as well

* update run name for noncurriculum

* lint

* dtype model eval

* return binary scoring

* set eval repeats to 3

* fix tests
2025-07-31 10:18:05 +02:00
joesharratt1229
4b60c32978
Curr exp (#487)
* began curr exp

* added holdout words

* updated config

* added context

* updated base curriculum

* updaed

* updated curriculum

* updated

* updated

* updated automatic flag

* updated ray trainer

* update
2025-07-25 20:38:47 +01:00
Oliver Stanley
1232a7d1e5
simplify training setup instructions (#454)
* simplify training setup instructions

* tweaks

* update cfgs

* readme update

* readme update
2025-06-06 09:51:29 +01:00
Oliver Stanley
add527ada1
update training dir with external eval details (#437)
* added games

* added llama 3b training conf

* update readme with details of external evals

* readme update

---------

Co-authored-by: joesharratt1229 <joesharratt1229@gmail.com>
2025-05-19 00:35:41 +02:00
joesharratt1229
73e3cb33a4
Added games training and evaluation configuration (#426)
* added games

* Update eval_games_composite.yaml

* Delete training/evaluations/eval_qwen_3b.yaml

* Add files via upload

* Delete training/evaluations/eval_algebraic_composite.yaml

* Delete training/evaluations/eval_algorithmic_composite.yaml

* Delete training/evaluations/eval_arithmetic_composite.yaml

* Delete training/evaluations/eval_cognition_composite.yaml

* Delete training/evaluations/eval_games_composite.yaml
2025-04-26 19:45:32 +01:00
joesharratt1229
d0ef136d5b
Feat/intragen experiments (#414)
* added curriculum

* readapted readme

* corrected small errors

* Delete eval/eval/r1/algorithmic/word_sorting.json

* removed redundant argument

* added spell

* removed duplicated fit

* changed config

* added composite changes

* added composite changes

* updated yaml

* added spell backward

* updated read me

* added qwen2.5

* added

* Add files via upload

* updated missing trainer func

* updated curr

* updated spell back

* updated correctness score func

* updated configs

* added local evals

* added updates

* updated datasets

* added fsdp to hf utility

* added algorithmic qwen 3b yaml

* updated read me

* updated configs

* added preappend token

* updated with thinking token

* updated test score board

* resolved comments

* added evaluation scripts

* removed results from pr

* added config

* added partial reward scoring

* added evaluation composites

* added training configs

* added games eval

* added rubriks cube

* resolved merge cinflicts

* added games config

* added latest eval configs

* updated strucutre

* Delete training/evaluations/eval_graphs_composite.yaml

---------

Co-authored-by: joesharratt1229 <joesharrat1229@gmail.com>
2025-04-16 08:04:52 +02:00
Oliver Stanley
224532f12a
first inter-domain generalisation experiments (#412)
* tweak len reward

* first inter-generalisation experiment config

* update inter algorithmic config

* default to empty config

* fix typo

* change config to match experiment script

* long prompt fixes

* algorithmic training config tweaks

* imports

* update algorithmic training cfgs

* first logic composite config

* fix dset name

* tweaks

* fix syllogisms dataset

* rm temp print

* initial algebra config

* algebra cfg tweaks

* add gc

* add initial games cfg

* rename games cfg

* fix dset name

* fix sokoban metadata

* remove boxnet

* games cfg tweak
2025-04-14 21:06:40 +01:00
joesharratt1229
43c739cb3e
Feat/curr adj (#394) 2025-04-02 06:39:14 +01:00
Zafir Stojanovski
c6663cdb81
fix(training): Prepend <think> token in format reward (#396)
* prepend think token in format reward

* pre commit + fix some default vals

* add checkpoint config
2025-03-28 09:45:17 +01:00
Oliver Stanley
eb69916c1b
initial verl training codebase (#389)
* fixes for latest verl
* composite dataset training experiment
* use stateful dataloaders to match verl changes
* training readme
* add formatting reward
* length reward impl
* standalone reasoning_gym config section
* curriculum learning, new length reward, more config
2025-03-20 15:04:57 +00:00