open-thought/reasoning-gym
1
0
Fork 0
mirror of https://github.com/open-thought/reasoning-gym.git synced 2026-04-25 17:10:51 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

cfg

Browse source
This commit is contained in:
Oliver 2025-04-22 21:15:08 +01:00
parent 1343bcf63e
commit 897e618bfa
1 changed files with 38 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

40
training/configs/external_generalisation/math_curriculum_qwen_7b.yaml
View file

@ -16,40 +16,76 @@ curriculum:
curricula:
complex_arithmetic:
weight: 1
attribute_levels:
'*': 0
intermediate_integration:
weight: 1
attribute_levels:
'*': 0
polynomial_equations:
weight: 1
attribute_levels:
'*': 0
polynomial_multiplication:
weight: 1
attribute_levels:
'*': 0
simple_equations:
weight: 1
attribute_levels:
'*': 0
simple_integration:
weight: 1
attribute_levels:
'*': 0
propositional_logic:
weight: 1
attribute_levels:
'*': 0
advanced_geometry:
weight: 1
attribute_levels:
'*': 0
simple_geometry:
weight: 1
attribute_levels:
'*': 0
basic_arithmetic:
weight: 1
attribute_levels:
'*': 0
bitwise_arithmetic:
weight: 1
attribute_levels:
'*': 0
chain_sum:
weight: 1
attribute_levels:
'*': 0
decimal_arithmetic:
weight: 1
attribute_levels:
'*': 0
decimal_chain_sum:
weight: 1
attribute_levels:
'*': 0
fraction_simplification:
weight: 1
attribute_levels:
'*': 0
gcd:
weight: 1
attribute_levels:
'*': 0
lcm:
weight: 1
attribute_levels:
'*': 0
prime_factorization:
weight: 1
attribute_levels:
'*': 0
reward:
use_accuracy: True
secondary_rewards:
@ -136,7 +172,7 @@ actor_rollout_ref:
enforce_eager: True
free_cache_engine: True
load_format: dummy_dtensor
tensor_model_parallel_size: 2
tensor_model_parallel_size: 1
max_num_batched_tokens: 12288
max_num_seqs: 1024
log_prob_micro_batch_size: null # will be deprecated, use log_prob_micro_batch_size_per_gpu
@ -172,7 +208,7 @@ trainer:
logger: [ 'console', 'wandb' ]
val_generations_to_log_to_wandb: 0
nnodes: 1
n_gpus_per_node: 2
n_gpus_per_node: 1
save_freq: 100
# auto: find the last ckpt to resume. If can't find, start from scratch
resume_mode: auto # or auto or resume_path if