forked from agentscope-ai/Trinity-RFT
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathsciworld.yaml
More file actions
59 lines (59 loc) · 1.32 KB
/
sciworld.yaml
File metadata and controls
59 lines (59 loc) · 1.32 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
global_config:
total_epochs: 20
batch_size: 4
algorithm_type: grpo
model:
model_path: '/PATH/TO/MODEL/CHECKPOINT/'
max_prompt_tokens: 4096
max_response_tokens: 16384
checkpoint_path: 'checkpoints/sciworld_RFT'
cluster:
node_num: 1
gpu_per_node: 8
buffer:
max_retry_times: 3
max_retry_interval: 1
explorer_input:
taskset:
name: sciworld
storage_type: file
path: 'scripts/data_prepare/sciworld_data'
format:
prompt_key: 'game_file'
rollout_args:
repeat_times: 8
temperature: 1.0
logprobs: 0
default_workflow_type: 'sciworld_workflow'
trainer_input:
experience_buffer:
name: sciworld_buffer
storage_type: queue
path: 'sqlite:///sciworld.db'
explorer:
engine_type: vllm_async
engine_num: 2
runner_num: 32
tensor_parallel_size: 2
enable_prefix_caching: false
enforce_eager: true
dtype: bfloat16
seed: 42
use_ray: false
backend: 'nccl'
max_pending_requests: 32
max_waiting_steps: 4
gpu_memory_utilization: 0.7
enable_chunked_prefill: true
synchronizer:
sync_method: 'nccl'
sync_interval: 8
sync_timeout: 1200
trainer:
trainer_type: 'verl'
trainer_config_path: 'examples/grpo_sciworld/train_sciworld.yaml'
save_interval: 10
monitor:
cache_root_dir: ""
project: "sciworld"
name: "sciworld_RFT"