Skip to content

Automation foundation model for tiny devices: 2-bit, 8-29 MB, tool calls, ASR, structured extraction and embeddings on phones, wearables, smart homes, robots, cars and microcontrollers.

use more representative post-epoch samples

HenryNdubuakub683be781fe5
diff --git a/src/train.py b/src/train.py
index 65ea052..8b3c8e3 100644
--- a/src/train.py
+++ b/src/train.py
@@ -891,10 +891,17 @@ def train(args):
_val_start = len(enc_inputs)
val_kept = val_data["kept_indices"]
- n_eval_samples = min(3, len(val_kept))
- step = max(1, len(val_kept) // n_eval_samples)
- sample_indices = [val_kept[i * step] for i in range(n_eval_samples)]
- eval_indices = np.array(sample_indices) + _val_start
+ n_eval_samples = 3
+ # Find samples with non-empty tool calls for representative eval
+ eval_indices = []
+ step = max(1, len(val_kept) // (n_eval_samples * 10))
+ for k in range(0, len(val_kept), step):
+ ds_idx = int(val_kept[k]) + _val_start
+ pair = load_example_with_audio(ds_idx)
+ if pair["answers"].strip() not in ("", "[]"):
+ eval_indices.append(ds_idx)
+ if len(eval_indices) >= n_eval_samples:
+ break
unified_samples = []
for i, ds_idx in enumerate(eval_indices):