findfile-xml-step240

findfile-xml-step240

WIP derivative checkpoint of EssentialAI/rnj-1-instruct.

RL-trained on file navigation task with XML tool-calling format.

Training Config

max_steps = 240
seq_len = 8192

[model]
name = "/home/ubuntu/models/hex-toolcall-base-v2"

[wandb]
project = "xml-tool-thinking"
name = "findfile-v2-from-hex140"

[orchestrator]
batch_size = 512
rollouts_per_example = 16

[orchestrator.sampling]
max_tokens = 512
temperature = 1.0

[[orchestrator.env]]
id = "findfile_xml"
args = { max_turns = 8, hint_chars = 40, noise_dirs = 3, max_read_chars = 1500, data_path = "/home/ubuntu/dataproc/consensus_pass_v2" }

[trainer.model]
ac = { freq = 1 }

[trainer.optim]
lr = 1e-6
max_norm = 0.001

[trainer.scheduler]
type = "linear"
warmup_steps = 5
decay_steps = 40
min_lr = 0

[inference.parallel]
tp = 4

[ckpt]