Upload Kairos checkpoint

README.md

@@ -1,60 +1,48 @@
 ---
 license: mit
-tags:
-- time-series
-- finance
-- kronos
-- kairos
-- crypto
-- btc
-- eth
+tags: [time-series, finance, kronos, kairos, crypto]
 library_name: pytorch
 pipeline_tag: time-series-forecasting
-base_model: NeoQuasar/Kronos-small
 ---
 
 # Shadowell/Kairos-small-crypto
 
-Fine-tuned **Kronos-small** on BTC/USDT + ETH/USDT 1-min K-lines (2024-01 ~ 2026-04)
-using [Kairos](https://github.com/Shadowell/Kairos).
+Fine-tuned **Kronos-small** on BTC/USDT + ETH/USDT 1-min K-lines (2024-01 ~ 2026-04) using **[Kairos](https://github.com/Shadowell/Kairos)**.
 Architecture = Kronos + exogenous bypass channel (32-d) + quantile return head.
 
-Training data came from the public Binance Vision spot mirror, so the 5 crypto-native
-exogenous features (funding_rate / funding_rate_z / oi_change / basis / btc_dominance)
-are padded to zero. The other 27 dims are real.
+This run keeps the original tokenizer [`NeoQuasar/Kronos-Tokenizer-base`](https://huggingface.co/NeoQuasar/Kronos-Tokenizer-base), matching the original `Kairos-small-crypto` training flow. Training data comes from the public Binance Vision spot mirror, so the 5 crypto-native exogenous features (`funding_rate` / `funding_rate_z` / `oi_change` / `basis` / `btc_dominance`) remain padded to zero; the other 27 dimensions are real.
 
-## Results on test set (2026-01-01 ~ 2026-04-16, ~304k 1-min bars)
+## Results on test set (2026-01-01 04:16:00 ~ 2026-04-16 23:30:00, 304,710 1-min bars)
 
-| horizon | model      | hit_rate | rank_ic | ICIR   |
-|---------|------------|---------:|--------:|-------:|
-| h1      | baseline   | 49.74%   | +0.022  | +0.297 |
-| h1      | finetuned  | 49.53%   | -0.012  | +0.029 |
-| h5      | baseline   | 49.72%   | -0.007  | +0.093 |
-| h5      | finetuned  | 50.51%   | +0.010  | +0.060 |
-| **h30** | **baseline**  | 50.98%   | +0.018 | +0.039 |
-| **h30** | **finetuned** | **51.68%** | **+0.050** | **+0.325** |
+| horizon | model | hit_rate | rank_ic | ICIR |
+|---|---|---:|---:|---:|
+| h1 | baseline | 50.58% | +0.001 | +0.184 |
+| h1 | finetuned | 49.53% | -0.012 | +0.029 |
+| h5 | baseline | 49.87% | -0.019 | -0.302 |
+| h5 | finetuned | 50.51% | +0.010 | +0.060 |
+| h30 | baseline | 49.04% | -0.026 | -0.140 |
+| h30 | finetuned | 51.68% | +0.050 | +0.325 |
 
 Baseline = original Kronos-small weights + randomly initialised exog / return head.
-h30 is the training target (`return_horizon=30` in the `crypto-1min` preset),
-so this is where the fine-tune is expected to help most -- rank-IC +184%, ICIR +743%.
-
-Full command log, pitfalls and a one-shot reproduction checklist are in
-[docs/CRYPTO_BTC_ETH_RUN.md](https://github.com/Shadowell/Kairos/blob/main/docs/CRYPTO_BTC_ETH_RUN.md).
+This rerun keeps the official `NeoQuasar/Kronos-Tokenizer-base`, matching the original `Kairos-small-crypto` flow.
+Training stopped at epoch 4; best val_ce = 2.4940.
 
 ## Usage
 
 ```python
-from kairos import KronosTokenizer
-from kairos.models import KronosWithExogenous
-
-tok   = KronosTokenizer.from_pretrained("NeoQuasar/Kronos-Tokenizer-base")
+from kairos import KronosTokenizer, KronosWithExogenous
+tok = KronosTokenizer.from_pretrained("NeoQuasar/Kronos-Tokenizer-base")
 model = KronosWithExogenous.from_pretrained("Shadowell/Kairos-small-crypto")
 ```
 
 ## Training config (preset `crypto-1min`)
 
 - lookback 256 min, predict 30 min
-- batch 50, OneCycleLR, early-stop patience 3 (stopped at epoch 4, best = epoch 1)
+- batch 50, OneCycleLR, early-stop patience 3
 - progressive unfreeze: only last transformer block + exog bypass + return head
+- tokenizer source = `{tok_repo}`
 - 32-d EXOG = 24 common + 8 crypto-market features
-- total wall time: 10 min 18 s on a single RTX 5090
+
+## Training recipe
+
+Full command log, backtest commands, pitfalls and the reproduction checklist are in [`docs/CRYPTO_BTC_ETH_RUN.md`](https://github.com/Shadowell/Kairos/blob/main/docs/CRYPTO_BTC_ETH_RUN.md).