---
license: apache-2.0
base_model: Qwen/Qwen3.5-35B-A3B
library_name: transformers
tags:
- qwen3_5
- moe
- medical
- sft
- deepmed
---

# Qwen3.5-35B-A3B-DeepMed-6task-SFT-epoch1

SFT of Qwen3.5-35B-A3B on DeepMed distilled trajectories (6 tasks) — **end of epoch 1**
(global step 210; epoch boundary at step 225).

- Base: `Qwen/Qwen3.5-35B-A3B`
- Data: 7,204 native-tool-call trajectories (distilled from 6 EHR tasks), filtered at 52K tokens
- Framework: verl + Megatron-Core (TP=2, EP=8, PP=1), bf16, full CPU optimizer offload
- Hyperparameters: lr 2e-5 (cosine), warmup 10 steps, weight decay 0.1, global batch 32
- Train loss: ~0.20 | Val loss: ~0.236 at this step

Companion checkpoint: [`Chtholly17/Qwen3.5-35B-A3B-DeepMed-6task-SFT-epoch2`](https://huggingface.co/Chtholly17/Qwen3.5-35B-A3B-DeepMed-6task-SFT-epoch2).