Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -11,4 +11,44 @@ RL-trained on file navigation task with XML tool-calling format.
|
|
| 11 |
|
| 12 |
## Training Config
|
| 13 |
|
|
|
|
|
|
|
|
|
|
| 14 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
|
| 12 |
## Training Config
|
| 13 |
|
| 14 |
+
```toml
|
| 15 |
+
max_steps = 240
|
| 16 |
+
seq_len = 8192
|
| 17 |
|
| 18 |
+
[model]
|
| 19 |
+
name = "/home/ubuntu/models/hex-toolcall-base-v2"
|
| 20 |
+
|
| 21 |
+
[wandb]
|
| 22 |
+
project = "xml-tool-thinking"
|
| 23 |
+
name = "findfile-v2-from-hex140"
|
| 24 |
+
|
| 25 |
+
[orchestrator]
|
| 26 |
+
batch_size = 512
|
| 27 |
+
rollouts_per_example = 16
|
| 28 |
+
|
| 29 |
+
[orchestrator.sampling]
|
| 30 |
+
max_tokens = 512
|
| 31 |
+
temperature = 1.0
|
| 32 |
+
|
| 33 |
+
[[orchestrator.env]]
|
| 34 |
+
id = "findfile_xml"
|
| 35 |
+
args = { max_turns = 8, hint_chars = 40, noise_dirs = 3, max_read_chars = 1500, data_path = "/home/ubuntu/dataproc/consensus_pass_v2" }
|
| 36 |
+
|
| 37 |
+
[trainer.model]
|
| 38 |
+
ac = { freq = 1 }
|
| 39 |
+
|
| 40 |
+
[trainer.optim]
|
| 41 |
+
lr = 1e-6
|
| 42 |
+
max_norm = 0.001
|
| 43 |
+
|
| 44 |
+
[trainer.scheduler]
|
| 45 |
+
type = "linear"
|
| 46 |
+
warmup_steps = 5
|
| 47 |
+
decay_steps = 40
|
| 48 |
+
min_lr = 0
|
| 49 |
+
|
| 50 |
+
[inference.parallel]
|
| 51 |
+
tp = 4
|
| 52 |
+
|
| 53 |
+
[ckpt]
|
| 54 |
+
```
|