Implementation:Shiyu coder Kronos Config Init
| Field | Value |
|---|---|
| implementation_name | Config_Init |
| type | API Doc |
| repository | https://github.com/shiyu-coder/Kronos |
| source_file | finetune/config.py:L3-131 |
| implements | Principle:Shiyu_coder_Kronos_Qlib_Experiment_Configuration |
| last_updated | 2026-02-09 14:00 GMT |
Summary
The Config class centralizes all experiment hyperparameters, data paths, time ranges, training settings, and backtesting parameters into a single configuration object with no constructor arguments.
Class
Config
API Signature
Config() -> Config
Import
from config import Config # run from finetune/ directory
Constructor
No constructor arguments. All fields are set as attributes in __init__.
Key Attributes
Data and Feature Parameters
| Attribute | Default | Description |
|---|---|---|
qlib_data_path |
"~/.qlib/qlib_data/cn_data" |
Path to Qlib CN data directory |
instrument |
"csi300" |
Instrument universe for data loading |
lookback_window |
90 |
Number of past time steps for input |
predict_window |
10 |
Number of future time steps for prediction |
max_context |
512 |
Maximum context length for the model |
feature_list |
['open', 'high', 'low', 'close', 'vol', 'amt'] |
Features used from raw data |
time_feature_list |
['minute', 'hour', 'weekday', 'day', 'month'] |
Time-based features to generate |
Time Ranges
| Attribute | Default | Description |
|---|---|---|
dataset_begin_time |
"2011-01-01" |
Overall data loading start |
dataset_end_time |
"2025-06-05" |
Overall data loading end |
train_time_range |
["2011-01-01", "2022-12-31"] |
Training split range |
val_time_range |
["2022-09-01", "2024-06-30"] |
Validation split range |
test_time_range |
["2024-04-01", "2025-06-05"] |
Test split range |
backtest_time_range |
["2024-07-01", "2025-06-05"] |
Backtesting time range |
Training Hyperparameters
| Attribute | Default | Description |
|---|---|---|
epochs |
30 |
Number of training epochs |
batch_size |
50 |
Batch size per GPU |
tokenizer_learning_rate |
2e-4 |
Learning rate for tokenizer finetuning |
predictor_learning_rate |
4e-5 |
Learning rate for predictor finetuning |
accumulation_steps |
1 |
Gradient accumulation steps |
adam_beta1 |
0.9 |
AdamW beta1 |
adam_beta2 |
0.95 |
AdamW beta2 |
adam_weight_decay |
0.1 |
AdamW weight decay |
clip |
5.0 |
Clipping value for normalized data |
n_train_iter |
2000 * batch_size |
Training samples per epoch |
n_val_iter |
400 * batch_size |
Validation samples per epoch |
seed |
100 |
Global random seed |
log_interval |
100 |
Log every N batches |
Paths
| Attribute | Default | Description |
|---|---|---|
save_path |
"./outputs/models" |
Base directory for model checkpoints |
dataset_path |
"./data/processed_datasets" |
Pickled dataset directory |
pretrained_tokenizer_path |
"path/to/your/Kronos-Tokenizer-base" |
Pretrained tokenizer location |
pretrained_predictor_path |
"path/to/your/Kronos-small" |
Pretrained predictor location |
finetuned_tokenizer_path |
(derived) | {save_path}/{tokenizer_save_folder_name}/checkpoints/best_model
|
finetuned_predictor_path |
(derived) | {save_path}/{predictor_save_folder_name}/checkpoints/best_model
|
backtest_result_path |
"./outputs/backtest_results" |
Backtest output directory |
Backtesting Parameters
| Attribute | Default | Description |
|---|---|---|
backtest_n_symbol_hold |
50 |
Number of symbols to hold in portfolio |
backtest_n_symbol_drop |
5 |
Number of symbols to drop per rebalance |
backtest_hold_thresh |
5 |
Minimum holding period for a stock |
inference_T |
0.6 |
Sampling temperature for inference |
inference_top_p |
0.9 |
Top-p (nucleus) sampling threshold |
inference_top_k |
0 |
Top-k sampling (0 = disabled) |
inference_sample_count |
5 |
Number of samples per inference |
backtest_batch_size |
1000 |
Batch size for inference during backtest |
backtest_benchmark |
(derived) | Benchmark index, set by _set_benchmark()
|
Experiment Logging
| Attribute | Default | Description |
|---|---|---|
use_comet |
True |
Enable/disable Comet ML logging |
comet_config |
{"api_key": ..., "project_name": ..., "workspace": ...} |
Comet ML connection settings |
comet_tag |
"finetune_demo" |
Tag for the Comet experiment |
comet_name |
"finetune_demo" |
Name for the Comet experiment |
Output
A Config instance with all attributes set. Can be converted to a dict via config.__dict__ for use in training functions.
Example Usage
from config import Config
config = Config()
# Access attributes directly
print(config.instrument) # "csi300"
print(config.epochs) # 30
print(config.lookback_window) # 90
# Convert to dict for training functions
config_dict = config.__dict__
Private Methods
_set_benchmark(instrument: str) -> str-- Maps instrument names to benchmark index codes (e.g.,"csi300"to"SH000300"). RaisesValueErrorif the instrument is not recognized.
Source Reference
File: finetune/config.py, lines 3-131.
Environment & Heuristic Links
Page Connections
Double-click a node to navigate. Hold to expand connections.
Principle
Implementation
Heuristic
Environment