v1.0.2

README.md

@@ -92,7 +92,7 @@ axolotl train examples/qwen3/qlora-fsdp-kd_qwen3-4b.yaml # 示例数据为作者
 
 
 ## SFT
-axolotl train examples/qwen3/qlora-fsdp-kd_qwen3-4b.yaml
+axolotl train examples/qwen3/qlora-fsdp_qwen3-4b.yaml
 
 
 ## RL
@@ -178,7 +178,7 @@ else:
 # Distil
 {'loss': 0.1297, 'grad_norm': 0.015708118677139282, 'learning_rate': 9.314053963669245e-07, 'epoch': 0.99}
 # SFT
-{'loss': 24.8723, 'grad_norm': 33.37716293334961, 'learning_rate': 2e-05, 'epoch': 0.09}
+{'loss': 2.3055, 'grad_norm': 4.508890151977539, 'learning_rate': 0.0, 'epoch': 0.01}
 # RL
 {'loss': 0.0044, 'grad_norm': 57.25, 'learning_rate': 3e-08, 'num_tokens': 57335.0, 'completions/mean_length': 348.53125, 'completions/min_length': 3.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.53125, 'completions/mean_terminated_length': 163.2666778564453, 'completions/min_terminated_length': 3.0, 'completions/max_terminated_length': 504.0, 'rewards/correctness_reward_func/mean': 0.1875, 'rewards/correctness_reward_func/std': 0.5922891497612, 'rewards/int_reward_func/mean': 0.046875, 'rewards/int_reward_func/std': 0.1480722874403, 'rewards/strict_format_reward_func/mean': 0.0, 'rewards/strict_format_reward_func/std': 0.0, 'rewards/soft_format_reward_func/mean': 0.0, 'rewards/soft_format_reward_func/std': 0.0, 'rewards/xmlcount_reward_func/mean': 0.06262499839067459, 'rewards/xmlcount_reward_func/std': 0.16834138333797455, 'reward': 0.2970000207424164, 'reward_std': 0.1547679901123047, 'kl': 0.005860495170054492, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.0, 'epoch': 0.01}
 ```

qwen3.sh

@@ -3,7 +3,7 @@ axolotl train examples/qwen3/qlora-fsdp-kd_qwen3-4b.yaml
 # axolotl train examples/qwen3/qlora-fsdp-kd_qwen3-4b_distilabel.yaml
 
 # SFT
-axolotl train examples/qwen3/qlora-fsdp-kd_qwen3-4b.yaml
+axolotl train examples/qwen3/qlora-fsdp_qwen3-4b.yaml
 
 # RL
 # rm -rf /root/.cache/*