qwen3-4b-tau2-oel-telecom-random74

qwen3-4b-tau2-oel-telecom-random74

Qwen3-4B-Instruct-2507 after one epoch of Online Experiential Learning (OEL) on the tau2-bench telecom domain.

What this is

OEL here is self-distillation, not RL — no reward signal is used. Teacher and student share weights; the only difference is that the teacher gets the task's memory appended to its system prompt. The student is fitted to the teacher's distribution with a token-level full KL loss, so the behaviour survives without the memory in the prompt.

Training

Evaluation

Not yet evaluated on the held-out test split.

Caveat

Across the runs in this series, OEL has not improved over the base model: differences sit within one standard error (SE ~0.06-0.08 at 40 tasks), and running a second epoch made things slightly worse rather than better. This matches what the method's authors report — putting the memories in the prompt at test time clearly helps, but distilling them into the weights recovers only part of that signal.

Gradient norms ran well above the 1.0 clip throughout, so the clip, not the learning rate, set the effective step size.

Published as a reproducible artifact, not as an improved model.

Usage

vllm serve MANGSEOK123/qwen3-4b-tau2-oel-telecom-random74 --enable-auto-tool-choice --tool-call-parser hermes