mirror of
https://github.com/open-jarvis/OpenJarvis.git
synced 2026-07-27 21:05:34 +00:00
Andrew Park
1a843d1ece
Tokenize the cleaned routing data (all-turns masking for ChatML),
full-parameter FSDP SFT (fsdp4/8 configs for Lambda 8xH100), and an eval
backend that drives the fine-tuned orchestrator as the scored model over
GAIA/MMLU-Pro/SuperGPQA. Includes split-making, Braintrust upload, eval
sample rendering, scorer fixes, and a boxed-tool-call salvage for SFT'd
checkpoints that emit their action inside \boxed{}.
26 lines
660 B
YAML
26 lines
660 B
YAML
compute_environment: LOCAL_MACHINE
|
|
distributed_type: FSDP
|
|
downcast_bf16: 'no'
|
|
machine_rank: 0
|
|
main_process_ip: null
|
|
main_process_port: null
|
|
main_training_function: main
|
|
mixed_precision: bf16
|
|
num_machines: 1
|
|
num_processes: 4
|
|
rdzv_backend: static
|
|
same_network: true
|
|
use_cpu: false
|
|
fsdp_config:
|
|
fsdp_version: 1
|
|
fsdp_auto_wrap_policy: SIZE_BASED_WRAP
|
|
fsdp_min_num_params: 100000000
|
|
fsdp_sharding_strategy: FULL_SHARD
|
|
fsdp_state_dict_type: FULL_STATE_DICT
|
|
fsdp_offload_params: false
|
|
fsdp_backward_prefetch: BACKWARD_PRE
|
|
fsdp_forward_prefetch: false
|
|
fsdp_use_orig_params: true
|
|
fsdp_cpu_ram_efficient_loading: true
|
|
fsdp_sync_module_states: true
|