Training in progress, step 100, checkpoint

.gitattributes

@@ -35,3 +35,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 tokenizer.json filter=lfs diff=lfs merge=lfs -text
 checkpoint-50/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+checkpoint-100/tokenizer.json filter=lfs diff=lfs merge=lfs -text

checkpoint-100/config.json

@@ -0,0 +1,29 @@
+{
+  "architectures": [
+    "Qwen2ForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151643,
+  "hidden_act": "silu",
+  "hidden_size": 1536,
+  "initializer_range": 0.02,
+  "intermediate_size": 8960,
+  "max_position_embeddings": 131072,
+  "max_window_layers": 21,
+  "model_type": "qwen2",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 2,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 10000,
+  "sliding_window": 4096,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.51.3",
+  "use_cache": false,
+  "use_mrope": false,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

checkpoint-100/generation_config.json

@@ -0,0 +1,9 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 151646,
+  "do_sample": true,
+  "eos_token_id": 151643,
+  "temperature": 0.6,
+  "top_p": 0.95,
+  "transformers_version": "4.51.3"
+}

checkpoint-100/global_step100/bf16_zero_pp_rank_0_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1670bd21d8cc03f32d0058e1f42c4a5042ff3bdfcb1cf5e9d6d27621ae550fbe
+size 5331274140

checkpoint-100/global_step100/bf16_zero_pp_rank_1_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:908a404bdf130542c101e1f908abd30cb1b205cd63e2ba4650f0c58d1179e4e8
+size 5331276572

checkpoint-100/global_step100/bf16_zero_pp_rank_2_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b12153d3157f6bbce7e6fbf3d7f4bdb411db9a5493bb62f824042b10d1772e34
+size 5331276892

checkpoint-100/global_step100/bf16_zero_pp_rank_3_mp_rank_00_optim_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1679bd7d909fa935484b80b16539d66bc89c72dc3c17b2cd6fbaf02f9b7dc01
+size 5331273884

checkpoint-100/global_step100/mp_rank_00_model_states.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69cbbae55733876e37fecae86b20bf7b09bbe36a1e60c1708f1f836c5e9028c6
+size 3554267640

checkpoint-100/latest

@@ -0,0 +1 @@
+global_step100

checkpoint-100/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4f456ef2898256a47b964478edbfa835c00a4c9ab5104dc3179ee0449604108
+size 3554214752

checkpoint-100/rng_state_0.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:888608037c15bb3968f3025caaa4aa0e62181f5e9f09e04aa188115c28fe8715
+size 14960

checkpoint-100/rng_state_1.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1f522d913077fd1366054cca2b40f4be35b08b338ddadab1cc1c30dd1820b71
+size 14960

checkpoint-100/rng_state_2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4744dca907a85a988cea74e47c179f1da08ca763ef98870d63ecddbd95a9b97e
+size 14960

checkpoint-100/rng_state_3.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:51ab4ed88c6dddf307cba8f9522e122811b83d254bff673bb15ccbfbeed289be
+size 14960

checkpoint-100/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b0d6ed3e119807e165d53a18d2ec22befd359c1465f6aeaa69d1d7eb1452246
+size 1064

checkpoint-100/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "<｜begin▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<｜end▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<｜end▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoint-100/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4256422650d141f228fe954acee98679da412984c29a569877eefd3af69315a
+size 11422959

checkpoint-100/tokenizer_config.json

@@ -0,0 +1,195 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "add_prefix_space": null,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<｜end▁of▁sentence｜>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<｜User｜>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151645": {
+      "content": "<｜Assistant｜>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151646": {
+      "content": "<｜begin▁of▁sentence｜>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|EOT|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151648": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151649": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "bos_token": "<｜begin▁of▁sentence｜>",
+  "chat_template": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='') %}{%- for message in messages %}{%- if message['role'] == 'system' %}{% set ns.system_prompt = message['content'] %}{%- endif %}{%- endfor %}{{bos_token}}{{ns.system_prompt}}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is none %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls']%}{%- if not ns.is_first %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin��>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- set ns.is_first = true -%}{%- else %}{{'\\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- endfor %}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is not none %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '</think>' in content %}{% set content = content.split('</think>')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'\\n<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜><think>\\n'}}{% endif %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<｜end▁of▁sentence｜>",
+  "extra_special_tokens": {},
+  "legacy": true,
+  "model_max_length": 16384,
+  "pad_token": "<｜end▁of▁sentence｜>",
+  "sp_model_kwargs": {},
+  "tokenizer_class": "LlamaTokenizerFast",
+  "unk_token": null,
+  "use_default_system_prompt": false
+}

checkpoint-100/trainer_state.json

@@ -0,0 +1,2534 @@
+{
+  "best_global_step": null,
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 0.11428571428571428,
+  "eval_steps": 500,
+  "global_step": 100,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.671875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1734.0,
+      "completions/mean_length": 1702.03125,
+      "completions/mean_terminated_length": 993.6190795898438,
+      "completions/min_length": 483.0,
+      "completions/min_terminated_length": 483.0,
+      "epoch": 0.001142857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.28377610445022583,
+      "learning_rate": 0.0,
+      "loss": -0.0,
+      "num_tokens": 118418.0,
+      "reward": -0.09800112247467041,
+      "reward_std": 0.3028089702129364,
+      "rewards/cosine_scaled_reward/mean": -0.09800112992525101,
+      "rewards/cosine_scaled_reward/std": 0.37953105568885803,
+      "step": 1
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1894.0,
+      "completions/mean_length": 1738.90625,
+      "completions/mean_terminated_length": 949.0,
+      "completions/min_length": 435.0,
+      "completions/min_terminated_length": 435.0,
+      "epoch": 0.002285714285714286,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24221572279930115,
+      "learning_rate": 2e-08,
+      "loss": -0.0,
+      "num_tokens": 239748.0,
+      "reward": 0.020556632429361343,
+      "reward_std": 0.3545936942100525,
+      "rewards/cosine_scaled_reward/mean": 0.020556632429361343,
+      "rewards/cosine_scaled_reward/std": 0.4492928683757782,
+      "step": 2
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.921875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1542.0,
+      "completions/mean_length": 1964.078125,
+      "completions/mean_terminated_length": 973.7999877929688,
+      "completions/min_length": 733.0,
+      "completions/min_terminated_length": 733.0,
+      "epoch": 0.0034285714285714284,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2472974807024002,
+      "learning_rate": 4e-08,
+      "loss": 0.0,
+      "num_tokens": 375921.0,
+      "reward": -0.20954538881778717,
+      "reward_std": 0.13813795149326324,
+      "rewards/cosine_scaled_reward/mean": -0.20954540371894836,
+      "rewards/cosine_scaled_reward/std": 0.16814909875392914,
+      "step": 3
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.484375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2010.0,
+      "completions/mean_length": 1555.6875,
+      "completions/mean_terminated_length": 1093.212158203125,
+      "completions/min_length": 502.0,
+      "completions/min_terminated_length": 502.0,
+      "epoch": 0.004571428571428572,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2868657112121582,
+      "learning_rate": 6e-08,
+      "loss": -0.0,
+      "num_tokens": 485293.0,
+      "reward": -0.12192361056804657,
+      "reward_std": 0.31710442900657654,
+      "rewards/cosine_scaled_reward/mean": -0.12192361056804657,
+      "rewards/cosine_scaled_reward/std": 0.35428565740585327,
+      "step": 4
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1758.0,
+      "completions/mean_length": 1958.5625,
+      "completions/mean_terminated_length": 1332.5,
+      "completions/min_length": 932.0,
+      "completions/min_terminated_length": 932.0,
+      "epoch": 0.005714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2460148185491562,
+      "learning_rate": 8e-08,
+      "loss": -0.0,
+      "num_tokens": 621457.0,
+      "reward": -0.21145480871200562,
+      "reward_std": 0.14890719950199127,
+      "rewards/cosine_scaled_reward/mean": -0.21145479381084442,
+      "rewards/cosine_scaled_reward/std": 0.20399661362171173,
+      "step": 5
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1670.0,
+      "completions/mean_length": 1908.375,
+      "completions/mean_terminated_length": 931.0,
+      "completions/min_length": 593.0,
+      "completions/min_terminated_length": 593.0,
+      "epoch": 0.006857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.26549720764160156,
+      "learning_rate": 1e-07,
+      "loss": -0.0,
+      "num_tokens": 755241.0,
+      "reward": -0.2408866286277771,
+      "reward_std": 0.16572487354278564,
+      "rewards/cosine_scaled_reward/mean": -0.2408866286277771,
+      "rewards/cosine_scaled_reward/std": 0.17492830753326416,
+      "step": 6
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.8125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1955.0,
+      "completions/mean_length": 1889.296875,
+      "completions/mean_terminated_length": 1201.5833740234375,
+      "completions/min_length": 396.0,
+      "completions/min_terminated_length": 396.0,
+      "epoch": 0.008,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23518230020999908,
+      "learning_rate": 1.2e-07,
+      "loss": 0.0,
+      "num_tokens": 886564.0,
+      "reward": -0.16087877750396729,
+      "reward_std": 0.24579641222953796,
+      "rewards/cosine_scaled_reward/mean": -0.16087877750396729,
+      "rewards/cosine_scaled_reward/std": 0.37339961528778076,
+      "step": 7
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1920.0,
+      "completions/mean_length": 1751.578125,
+      "completions/mean_terminated_length": 994.0555419921875,
+      "completions/min_length": 330.0,
+      "completions/min_terminated_length": 330.0,
+      "epoch": 0.009142857142857144,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2354528158903122,
+      "learning_rate": 1.4e-07,
+      "loss": 0.0,
+      "num_tokens": 1009081.0,
+      "reward": -0.023812226951122284,
+      "reward_std": 0.2823081314563751,
+      "rewards/cosine_scaled_reward/mean": -0.02381223440170288,
+      "rewards/cosine_scaled_reward/std": 0.4484662115573883,
+      "step": 8
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.9375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1889.0,
+      "completions/mean_length": 2000.59375,
+      "completions/mean_terminated_length": 1289.5,
+      "completions/min_length": 903.0,
+      "completions/min_terminated_length": 903.0,
+      "epoch": 0.010285714285714285,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24302220344543457,
+      "learning_rate": 1.6e-07,
+      "loss": 0.0,
+      "num_tokens": 1148575.0,
+      "reward": -0.2453702688217163,
+      "reward_std": 0.18811637163162231,
+      "rewards/cosine_scaled_reward/mean": -0.2453702688217163,
+      "rewards/cosine_scaled_reward/std": 0.22203005850315094,
+      "step": 9
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.703125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1739.0,
+      "completions/mean_length": 1701.140625,
+      "completions/mean_terminated_length": 879.631591796875,
+      "completions/min_length": 484.0,
+      "completions/min_terminated_length": 484.0,
+      "epoch": 0.011428571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.25642141699790955,
+      "learning_rate": 1.8e-07,
+      "loss": -0.0,
+      "num_tokens": 1268280.0,
+      "reward": -0.15177705883979797,
+      "reward_std": 0.2125300019979477,
+      "rewards/cosine_scaled_reward/mean": -0.15177705883979797,
+      "rewards/cosine_scaled_reward/std": 0.3240113854408264,
+      "step": 10
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.890625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1683.0,
+      "completions/mean_length": 1950.609375,
+      "completions/mean_terminated_length": 1157.571533203125,
+      "completions/min_length": 584.0,
+      "completions/min_terminated_length": 584.0,
+      "epoch": 0.012571428571428572,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24372951686382294,
+      "learning_rate": 2e-07,
+      "loss": 0.0,
+      "num_tokens": 1404791.0,
+      "reward": -0.23502977192401886,
+      "reward_std": 0.18896539509296417,
+      "rewards/cosine_scaled_reward/mean": -0.23502977192401886,
+      "rewards/cosine_scaled_reward/std": 0.24224351346492767,
+      "step": 11
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.640625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1944.0,
+      "completions/mean_length": 1751.03125,
+      "completions/mean_terminated_length": 1221.6522216796875,
+      "completions/min_length": 489.0,
+      "completions/min_terminated_length": 489.0,
+      "epoch": 0.013714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.28422027826309204,
+      "learning_rate": 2.1999999999999998e-07,
+      "loss": -0.0,
+      "num_tokens": 1527801.0,
+      "reward": -0.14280016720294952,
+      "reward_std": 0.32843896746635437,
+      "rewards/cosine_scaled_reward/mean": -0.14280015230178833,
+      "rewards/cosine_scaled_reward/std": 0.41895967721939087,
+      "step": 12
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1793.0,
+      "completions/mean_length": 1834.453125,
+      "completions/mean_terminated_length": 1193.8125,
+      "completions/min_length": 783.0,
+      "completions/min_terminated_length": 783.0,
+      "epoch": 0.014857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24033738672733307,
+      "learning_rate": 2.4e-07,
+      "loss": 0.0,
+      "num_tokens": 1656246.0,
+      "reward": -0.17057427763938904,
+      "reward_std": 0.24429668486118317,
+      "rewards/cosine_scaled_reward/mean": -0.17057427763938904,
+      "rewards/cosine_scaled_reward/std": 0.27816399931907654,
+      "step": 13
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1474.0,
+      "completions/mean_length": 1800.65625,
+      "completions/mean_terminated_length": 1116.823486328125,
+      "completions/min_length": 495.0,
+      "completions/min_terminated_length": 495.0,
+      "epoch": 0.016,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2312558889389038,
+      "learning_rate": 2.6e-07,
+      "loss": 0.0,
+      "num_tokens": 1782096.0,
+      "reward": -0.11817245185375214,
+      "reward_std": 0.24491220712661743,
+      "rewards/cosine_scaled_reward/mean": -0.11817245930433273,
+      "rewards/cosine_scaled_reward/std": 0.3942086696624756,
+      "step": 14
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1636.0,
+      "completions/mean_length": 1692.828125,
+      "completions/mean_terminated_length": 785.1666870117188,
+      "completions/min_length": 438.0,
+      "completions/min_terminated_length": 438.0,
+      "epoch": 0.017142857142857144,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2563658654689789,
+      "learning_rate": 2.8e-07,
+      "loss": -0.0,
+      "num_tokens": 1901357.0,
+      "reward": -0.027107469737529755,
+      "reward_std": 0.1853453516960144,
+      "rewards/cosine_scaled_reward/mean": -0.027107462286949158,
+      "rewards/cosine_scaled_reward/std": 0.4734213352203369,
+      "step": 15
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 1.0,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 0.0,
+      "completions/mean_length": 2048.0,
+      "completions/mean_terminated_length": 0.0,
+      "completions/min_length": 2048.0,
+      "completions/min_terminated_length": 0.0,
+      "epoch": 0.018285714285714287,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24149107933044434,
+      "learning_rate": 3e-07,
+      "loss": -0.0,
+      "num_tokens": 2042869.0,
+      "reward": -0.2542623281478882,
+      "reward_std": 0.14302438497543335,
+      "rewards/cosine_scaled_reward/mean": -0.2542623281478882,
+      "rewards/cosine_scaled_reward/std": 0.160969540476799,
+      "step": 16
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.578125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1827.0,
+      "completions/mean_length": 1548.75,
+      "completions/mean_terminated_length": 864.5925903320312,
+      "completions/min_length": 357.0,
+      "completions/min_terminated_length": 357.0,
+      "epoch": 0.019428571428571427,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.31088724732398987,
+      "learning_rate": 3.2e-07,
+      "loss": 0.0,
+      "num_tokens": 2152509.0,
+      "reward": -0.12113451957702637,
+      "reward_std": 0.284165620803833,
+      "rewards/cosine_scaled_reward/mean": -0.12113452702760696,
+      "rewards/cosine_scaled_reward/std": 0.4259316623210907,
+      "step": 17
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1899.0,
+      "completions/mean_length": 1793.03125,
+      "completions/mean_terminated_length": 1028.125,
+      "completions/min_length": 531.0,
+      "completions/min_terminated_length": 531.0,
+      "epoch": 0.02057142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2451843023300171,
+      "learning_rate": 3.4000000000000003e-07,
+      "loss": 0.0,
+      "num_tokens": 2277639.0,
+      "reward": -0.18317042291164398,
+      "reward_std": 0.20634235441684723,
+      "rewards/cosine_scaled_reward/mean": -0.18317043781280518,
+      "rewards/cosine_scaled_reward/std": 0.27781662344932556,
+      "step": 18
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.703125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1633.0,
+      "completions/mean_length": 1735.984375,
+      "completions/mean_terminated_length": 997.0,
+      "completions/min_length": 462.0,
+      "completions/min_terminated_length": 462.0,
+      "epoch": 0.021714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24677637219429016,
+      "learning_rate": 3.6e-07,
+      "loss": 0.0,
+      "num_tokens": 2399998.0,
+      "reward": -0.04996331408619881,
+      "reward_std": 0.2841629385948181,
+      "rewards/cosine_scaled_reward/mean": -0.04996330291032791,
+      "rewards/cosine_scaled_reward/std": 0.4186851680278778,
+      "step": 19
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.640625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1643.0,
+      "completions/mean_length": 1614.890625,
+      "completions/mean_terminated_length": 842.8261108398438,
+      "completions/min_length": 411.0,
+      "completions/min_terminated_length": 411.0,
+      "epoch": 0.022857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2543003559112549,
+      "learning_rate": 3.7999999999999996e-07,
+      "loss": -0.0,
+      "num_tokens": 2514703.0,
+      "reward": -0.09282197058200836,
+      "reward_std": 0.2568933367729187,
+      "rewards/cosine_scaled_reward/mean": -0.09282197058200836,
+      "rewards/cosine_scaled_reward/std": 0.4104878604412079,
+      "step": 20
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1988.0,
+      "completions/mean_length": 1786.734375,
+      "completions/mean_terminated_length": 1119.0555419921875,
+      "completions/min_length": 348.0,
+      "completions/min_terminated_length": 348.0,
+      "epoch": 0.024,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.3147278130054474,
+      "learning_rate": 4e-07,
+      "loss": 0.0,
+      "num_tokens": 2639862.0,
+      "reward": -0.16029146313667297,
+      "reward_std": 0.2322564721107483,
+      "rewards/cosine_scaled_reward/mean": -0.16029146313667297,
+      "rewards/cosine_scaled_reward/std": 0.36191171407699585,
+      "step": 21
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.40625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1833.0,
+      "completions/mean_length": 1300.484375,
+      "completions/mean_terminated_length": 789.0263061523438,
+      "completions/min_length": 287.0,
+      "completions/min_terminated_length": 287.0,
+      "epoch": 0.025142857142857144,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.32522445917129517,
+      "learning_rate": 4.1999999999999995e-07,
+      "loss": 0.0,
+      "num_tokens": 2732109.0,
+      "reward": 0.0033364146947860718,
+      "reward_std": 0.18878400325775146,
+      "rewards/cosine_scaled_reward/mean": 0.0033364109694957733,
+      "rewards/cosine_scaled_reward/std": 0.45390966534614563,
+      "step": 22
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.59375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1920.0,
+      "completions/mean_length": 1641.03125,
+      "completions/mean_terminated_length": 1046.2308349609375,
+      "completions/min_length": 422.0,
+      "completions/min_terminated_length": 422.0,
+      "epoch": 0.026285714285714287,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.28244850039482117,
+      "learning_rate": 4.3999999999999997e-07,
+      "loss": 0.0,
+      "num_tokens": 2847927.0,
+      "reward": -0.21077856421470642,
+      "reward_std": 0.24399788677692413,
+      "rewards/cosine_scaled_reward/mean": -0.21077856421470642,
+      "rewards/cosine_scaled_reward/std": 0.2925592362880707,
+      "step": 23
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1965.0,
+      "completions/mean_length": 1789.59375,
+      "completions/mean_terminated_length": 1129.2222900390625,
+      "completions/min_length": 560.0,
+      "completions/min_terminated_length": 560.0,
+      "epoch": 0.027428571428571427,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24896308779716492,
+      "learning_rate": 4.6e-07,
+      "loss": -0.0,
+      "num_tokens": 2973389.0,
+      "reward": -0.1665852814912796,
+      "reward_std": 0.307574987411499,
+      "rewards/cosine_scaled_reward/mean": -0.1665852665901184,
+      "rewards/cosine_scaled_reward/std": 0.4072873294353485,
+      "step": 24
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.65625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1851.0,
+      "completions/mean_length": 1696.40625,
+      "completions/mean_terminated_length": 1025.181884765625,
+      "completions/min_length": 434.0,
+      "completions/min_terminated_length": 434.0,
+      "epoch": 0.02857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.262716144323349,
+      "learning_rate": 4.8e-07,
+      "loss": 0.0,
+      "num_tokens": 3092255.0,
+      "reward": -0.14361324906349182,
+      "reward_std": 0.3466429114341736,
+      "rewards/cosine_scaled_reward/mean": -0.14361326396465302,
+      "rewards/cosine_scaled_reward/std": 0.3933021128177643,
+      "step": 25
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1975.0,
+      "completions/mean_length": 1973.046875,
+      "completions/mean_terminated_length": 1448.375,
+      "completions/min_length": 1035.0,
+      "completions/min_terminated_length": 1035.0,
+      "epoch": 0.029714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2365841567516327,
+      "learning_rate": 5e-07,
+      "loss": -0.0,
+      "num_tokens": 3229162.0,
+      "reward": -0.050574399530887604,
+      "reward_std": 0.22459164261817932,
+      "rewards/cosine_scaled_reward/mean": -0.050574399530887604,
+      "rewards/cosine_scaled_reward/std": 0.37290775775909424,
+      "step": 26
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.796875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1982.0,
+      "completions/mean_length": 1878.53125,
+      "completions/mean_terminated_length": 1213.6923828125,
+      "completions/min_length": 498.0,
+      "completions/min_terminated_length": 498.0,
+      "epoch": 0.030857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2821083068847656,
+      "learning_rate": 5.2e-07,
+      "loss": 0.0,
+      "num_tokens": 3359676.0,
+      "reward": -0.13096781075000763,
+      "reward_std": 0.26249831914901733,
+      "rewards/cosine_scaled_reward/mean": -0.13096781075000763,
+      "rewards/cosine_scaled_reward/std": 0.3478032350540161,
+      "step": 27
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.78125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1933.0,
+      "completions/mean_length": 1827.453125,
+      "completions/mean_terminated_length": 1039.7857666015625,
+      "completions/min_length": 419.0,
+      "completions/min_terminated_length": 419.0,
+      "epoch": 0.032,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2539210915565491,
+      "learning_rate": 5.4e-07,
+      "loss": 0.0,
+      "num_tokens": 3486969.0,
+      "reward": -0.11822876334190369,
+      "reward_std": 0.2370690554380417,
+      "rewards/cosine_scaled_reward/mean": -0.11822875589132309,
+      "rewards/cosine_scaled_reward/std": 0.4236762225627899,
+      "step": 28
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.9375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2047.0,
+      "completions/mean_length": 2020.5,
+      "completions/mean_terminated_length": 1608.0,
+      "completions/min_length": 887.0,
+      "completions/min_terminated_length": 887.0,
+      "epoch": 0.03314285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23259545862674713,
+      "learning_rate": 5.6e-07,
+      "loss": -0.0,
+      "num_tokens": 3626753.0,
+      "reward": -0.20220182836055756,
+      "reward_std": 0.15910759568214417,
+      "rewards/cosine_scaled_reward/mean": -0.20220182836055756,
+      "rewards/cosine_scaled_reward/std": 0.20781411230564117,
+      "step": 29
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.828125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1662.0,
+      "completions/mean_length": 1903.703125,
+      "completions/mean_terminated_length": 1208.45458984375,
+      "completions/min_length": 961.0,
+      "completions/min_terminated_length": 961.0,
+      "epoch": 0.03428571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24027252197265625,
+      "learning_rate": 5.8e-07,
+      "loss": 0.0,
+      "num_tokens": 3759126.0,
+      "reward": -0.19193249940872192,
+      "reward_std": 0.24584847688674927,
+      "rewards/cosine_scaled_reward/mean": -0.19193249940872192,
+      "rewards/cosine_scaled_reward/std": 0.28378522396087646,
+      "step": 30
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.796875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1922.0,
+      "completions/mean_length": 1847.34375,
+      "completions/mean_terminated_length": 1060.1539306640625,
+      "completions/min_length": 311.0,
+      "completions/min_terminated_length": 311.0,
+      "epoch": 0.03542857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2703397274017334,
+      "learning_rate": 6e-07,
+      "loss": -0.0,
+      "num_tokens": 3887852.0,
+      "reward": -0.25379180908203125,
+      "reward_std": 0.24661941826343536,
+      "rewards/cosine_scaled_reward/mean": -0.25379180908203125,
+      "rewards/cosine_scaled_reward/std": 0.29188498854637146,
+      "step": 31
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.828125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1995.0,
+      "completions/mean_length": 1950.3125,
+      "completions/mean_terminated_length": 1479.6363525390625,
+      "completions/min_length": 766.0,
+      "completions/min_terminated_length": 766.0,
+      "epoch": 0.036571428571428574,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21763876080513,
+      "learning_rate": 6.2e-07,
+      "loss": -0.0,
+      "num_tokens": 4023024.0,
+      "reward": -0.16017228364944458,
+      "reward_std": 0.2255343496799469,
+      "rewards/cosine_scaled_reward/mean": -0.16017228364944458,
+      "rewards/cosine_scaled_reward/std": 0.3709539771080017,
+      "step": 32
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1957.0,
+      "completions/mean_length": 1996.28125,
+      "completions/mean_terminated_length": 1634.25,
+      "completions/min_length": 1237.0,
+      "completions/min_terminated_length": 1237.0,
+      "epoch": 0.037714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22758260369300842,
+      "learning_rate": 6.4e-07,
+      "loss": -0.0,
+      "num_tokens": 4162002.0,
+      "reward": -0.20318198204040527,
+      "reward_std": 0.18396919965744019,
+      "rewards/cosine_scaled_reward/mean": -0.20318198204040527,
+      "rewards/cosine_scaled_reward/std": 0.34913352131843567,
+      "step": 33
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.578125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1850.0,
+      "completions/mean_length": 1703.265625,
+      "completions/mean_terminated_length": 1230.851806640625,
+      "completions/min_length": 651.0,
+      "completions/min_terminated_length": 651.0,
+      "epoch": 0.038857142857142854,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.31658875942230225,
+      "learning_rate": 6.6e-07,
+      "loss": -0.0,
+      "num_tokens": 4280563.0,
+      "reward": -0.05977274850010872,
+      "reward_std": 0.30437377095222473,
+      "rewards/cosine_scaled_reward/mean": -0.059772733598947525,
+      "rewards/cosine_scaled_reward/std": 0.4424094259738922,
+      "step": 34
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.8125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1846.0,
+      "completions/mean_length": 1807.546875,
+      "completions/mean_terminated_length": 765.5833740234375,
+      "completions/min_length": 419.0,
+      "completions/min_terminated_length": 419.0,
+      "epoch": 0.04,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2792847156524658,
+      "learning_rate": 6.800000000000001e-07,
+      "loss": -0.0,
+      "num_tokens": 4407742.0,
+      "reward": -0.18658886849880219,
+      "reward_std": 0.2910658121109009,
+      "rewards/cosine_scaled_reward/mean": -0.18658888339996338,
+      "rewards/cosine_scaled_reward/std": 0.34802255034446716,
+      "step": 35
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.921875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1912.0,
+      "completions/mean_length": 1995.65625,
+      "completions/mean_terminated_length": 1378.0,
+      "completions/min_length": 1090.0,
+      "completions/min_terminated_length": 1090.0,
+      "epoch": 0.04114285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23547738790512085,
+      "learning_rate": 7e-07,
+      "loss": 0.0,
+      "num_tokens": 4546576.0,
+      "reward": -0.23918019235134125,
+      "reward_std": 0.19598917663097382,
+      "rewards/cosine_scaled_reward/mean": -0.23918019235134125,
+      "rewards/cosine_scaled_reward/std": 0.2425125539302826,
+      "step": 36
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.90625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2015.0,
+      "completions/mean_length": 1994.75,
+      "completions/mean_terminated_length": 1480.0,
+      "completions/min_length": 545.0,
+      "completions/min_terminated_length": 545.0,
+      "epoch": 0.04228571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22962674498558044,
+      "learning_rate": 7.2e-07,
+      "loss": -0.0,
+      "num_tokens": 4685264.0,
+      "reward": -0.25335729122161865,
+      "reward_std": 0.15323391556739807,
+      "rewards/cosine_scaled_reward/mean": -0.25335729122161865,
+      "rewards/cosine_scaled_reward/std": 0.17556406557559967,
+      "step": 37
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.890625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1934.0,
+      "completions/mean_length": 1957.484375,
+      "completions/mean_terminated_length": 1220.4285888671875,
+      "completions/min_length": 965.0,
+      "completions/min_terminated_length": 965.0,
+      "epoch": 0.04342857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24781912565231323,
+      "learning_rate": 7.4e-07,
+      "loss": -0.0,
+      "num_tokens": 4822255.0,
+      "reward": -0.13536512851715088,
+      "reward_std": 0.19208545982837677,
+      "rewards/cosine_scaled_reward/mean": -0.13536511361598969,
+      "rewards/cosine_scaled_reward/std": 0.30052343010902405,
+      "step": 38
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1824.0,
+      "completions/mean_length": 1744.421875,
+      "completions/mean_terminated_length": 833.6875,
+      "completions/min_length": 317.0,
+      "completions/min_terminated_length": 317.0,
+      "epoch": 0.044571428571428574,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2562144994735718,
+      "learning_rate": 7.599999999999999e-07,
+      "loss": -0.0,
+      "num_tokens": 4944682.0,
+      "reward": -0.041110455989837646,
+      "reward_std": 0.21381449699401855,
+      "rewards/cosine_scaled_reward/mean": -0.04111045226454735,
+      "rewards/cosine_scaled_reward/std": 0.35980772972106934,
+      "step": 39
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2043.0,
+      "completions/mean_length": 1774.359375,
+      "completions/mean_terminated_length": 1017.8235473632812,
+      "completions/min_length": 445.0,
+      "completions/min_terminated_length": 445.0,
+      "epoch": 0.045714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.25478634238243103,
+      "learning_rate": 7.799999999999999e-07,
+      "loss": 0.0,
+      "num_tokens": 5068313.0,
+      "reward": -0.12165145576000214,
+      "reward_std": 0.17204006016254425,
+      "rewards/cosine_scaled_reward/mean": -0.12165144830942154,
+      "rewards/cosine_scaled_reward/std": 0.4099982678890228,
+      "step": 40
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.640625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1991.0,
+      "completions/mean_length": 1814.375,
+      "completions/mean_terminated_length": 1397.9130859375,
+      "completions/min_length": 968.0,
+      "completions/min_terminated_length": 968.0,
+      "epoch": 0.046857142857142854,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21750310063362122,
+      "learning_rate": 8e-07,
+      "loss": 0.0,
+      "num_tokens": 5195585.0,
+      "reward": -0.25668060779571533,
+      "reward_std": 0.2832298278808594,
+      "rewards/cosine_scaled_reward/mean": -0.25668060779571533,
+      "rewards/cosine_scaled_reward/std": 0.3347759544849396,
+      "step": 41
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.765625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1764.0,
+      "completions/mean_length": 1714.59375,
+      "completions/mean_terminated_length": 625.4666748046875,
+      "completions/min_length": 186.0,
+      "completions/min_terminated_length": 186.0,
+      "epoch": 0.048,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.34486907720565796,
+      "learning_rate": 8.199999999999999e-07,
+      "loss": -0.0,
+      "num_tokens": 5315679.0,
+      "reward": -0.2253742218017578,
+      "reward_std": 0.1778060495853424,
+      "rewards/cosine_scaled_reward/mean": -0.22537420690059662,
+      "rewards/cosine_scaled_reward/std": 0.19647939503192902,
+      "step": 42
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.828125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1638.0,
+      "completions/mean_length": 1863.78125,
+      "completions/mean_terminated_length": 976.1818237304688,
+      "completions/min_length": 669.0,
+      "completions/min_terminated_length": 669.0,
+      "epoch": 0.04914285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23907455801963806,
+      "learning_rate": 8.399999999999999e-07,
+      "loss": 0.0,
+      "num_tokens": 5446577.0,
+      "reward": -0.1142776757478714,
+      "reward_std": 0.21804723143577576,
+      "rewards/cosine_scaled_reward/mean": -0.1142776757478714,
+      "rewards/cosine_scaled_reward/std": 0.3637608587741852,
+      "step": 43
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1920.0,
+      "completions/mean_length": 1771.125,
+      "completions/mean_terminated_length": 940.5,
+      "completions/min_length": 344.0,
+      "completions/min_terminated_length": 344.0,
+      "epoch": 0.05028571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2888188362121582,
+      "learning_rate": 8.599999999999999e-07,
+      "loss": 0.0,
+      "num_tokens": 5570625.0,
+      "reward": -0.11845305562019348,
+      "reward_std": 0.2729855477809906,
+      "rewards/cosine_scaled_reward/mean": -0.11845306307077408,
+      "rewards/cosine_scaled_reward/std": 0.4279690086841583,
+      "step": 44
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.96875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1296.0,
+      "completions/mean_length": 2020.859375,
+      "completions/mean_terminated_length": 1179.5,
+      "completions/min_length": 1063.0,
+      "completions/min_terminated_length": 1063.0,
+      "epoch": 0.05142857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2232045829296112,
+      "learning_rate": 8.799999999999999e-07,
+      "loss": 0.0,
+      "num_tokens": 5711616.0,
+      "reward": -0.1830526441335678,
+      "reward_std": 0.20074567198753357,
+      "rewards/cosine_scaled_reward/mean": -0.1830526441335678,
+      "rewards/cosine_scaled_reward/std": 0.3221423327922821,
+      "step": 45
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.828125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1121.0,
+      "completions/mean_length": 1843.328125,
+      "completions/mean_terminated_length": 857.1818237304688,
+      "completions/min_length": 608.0,
+      "completions/min_terminated_length": 608.0,
+      "epoch": 0.052571428571428575,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2569328844547272,
+      "learning_rate": 9e-07,
+      "loss": 0.0,
+      "num_tokens": 5840757.0,
+      "reward": -0.21247822046279907,
+      "reward_std": 0.17188501358032227,
+      "rewards/cosine_scaled_reward/mean": -0.21247822046279907,
+      "rewards/cosine_scaled_reward/std": 0.183182492852211,
+      "step": 46
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2047.0,
+      "completions/mean_length": 1772.984375,
+      "completions/mean_terminated_length": 1012.6470336914062,
+      "completions/min_length": 461.0,
+      "completions/min_terminated_length": 461.0,
+      "epoch": 0.053714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2800576090812683,
+      "learning_rate": 9.2e-07,
+      "loss": -0.0,
+      "num_tokens": 5964628.0,
+      "reward": -0.1755329668521881,
+      "reward_std": 0.19662824273109436,
+      "rewards/cosine_scaled_reward/mean": -0.1755329668521881,
+      "rewards/cosine_scaled_reward/std": 0.3987559974193573,
+      "step": 47
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1949.0,
+      "completions/mean_length": 1787.046875,
+      "completions/mean_terminated_length": 1120.1666259765625,
+      "completions/min_length": 630.0,
+      "completions/min_terminated_length": 630.0,
+      "epoch": 0.054857142857142854,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2499135434627533,
+      "learning_rate": 9.399999999999999e-07,
+      "loss": -0.0,
+      "num_tokens": 6089543.0,
+      "reward": -0.07469595968723297,
+      "reward_std": 0.2802818715572357,
+      "rewards/cosine_scaled_reward/mean": -0.07469595968723297,
+      "rewards/cosine_scaled_reward/std": 0.39331451058387756,
+      "step": 48
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.578125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1818.0,
+      "completions/mean_length": 1611.65625,
+      "completions/mean_terminated_length": 1013.7037353515625,
+      "completions/min_length": 298.0,
+      "completions/min_terminated_length": 298.0,
+      "epoch": 0.056,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2976716160774231,
+      "learning_rate": 9.6e-07,
+      "loss": -0.0,
+      "num_tokens": 6202753.0,
+      "reward": -0.14219576120376587,
+      "reward_std": 0.3252427875995636,
+      "rewards/cosine_scaled_reward/mean": -0.14219576120376587,
+      "rewards/cosine_scaled_reward/std": 0.41946855187416077,
+      "step": 49
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.828125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1916.0,
+      "completions/mean_length": 1826.90625,
+      "completions/mean_terminated_length": 761.6364135742188,
+      "completions/min_length": 341.0,
+      "completions/min_terminated_length": 341.0,
+      "epoch": 0.05714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2344626933336258,
+      "learning_rate": 9.8e-07,
+      "loss": -0.0,
+      "num_tokens": 6330491.0,
+      "reward": -0.098542720079422,
+      "reward_std": 0.20483215153217316,
+      "rewards/cosine_scaled_reward/mean": -0.0985427126288414,
+      "rewards/cosine_scaled_reward/std": 0.396296888589859,
+      "step": 50
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.53125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2034.0,
+      "completions/mean_length": 1520.1875,
+      "completions/mean_terminated_length": 922.0000610351562,
+      "completions/min_length": 450.0,
+      "completions/min_terminated_length": 450.0,
+      "epoch": 0.05828571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.30348992347717285,
+      "learning_rate": 1e-06,
+      "loss": 0.0,
+      "num_tokens": 6437991.0,
+      "reward": -0.12996003031730652,
+      "reward_std": 0.2803010940551758,
+      "rewards/cosine_scaled_reward/mean": -0.12996003031730652,
+      "rewards/cosine_scaled_reward/std": 0.3464147746562958,
+      "step": 51
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1930.0,
+      "completions/mean_length": 1726.71875,
+      "completions/mean_terminated_length": 838.4705810546875,
+      "completions/min_length": 315.0,
+      "completions/min_terminated_length": 315.0,
+      "epoch": 0.05942857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2905585765838623,
+      "learning_rate": 9.999890338174275e-07,
+      "loss": -0.0,
+      "num_tokens": 6559853.0,
+      "reward": -0.2443142831325531,
+      "reward_std": 0.21010473370552063,
+      "rewards/cosine_scaled_reward/mean": -0.2443142831325531,
+      "rewards/cosine_scaled_reward/std": 0.32864055037498474,
+      "step": 52
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1585.0,
+      "completions/mean_length": 1757.015625,
+      "completions/mean_terminated_length": 952.5294189453125,
+      "completions/min_length": 463.0,
+      "completions/min_terminated_length": 463.0,
+      "epoch": 0.060571428571428575,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2699633538722992,
+      "learning_rate": 9.999561358041868e-07,
+      "loss": 0.0,
+      "num_tokens": 6683134.0,
+      "reward": -0.18116676807403564,
+      "reward_std": 0.2308851182460785,
+      "rewards/cosine_scaled_reward/mean": -0.18116676807403564,
+      "rewards/cosine_scaled_reward/std": 0.27486056089401245,
+      "step": 53
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.765625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2018.0,
+      "completions/mean_length": 1850.65625,
+      "completions/mean_terminated_length": 1206.0001220703125,
+      "completions/min_length": 695.0,
+      "completions/min_terminated_length": 695.0,
+      "epoch": 0.061714285714285715,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23383355140686035,
+      "learning_rate": 9.999013075636804e-07,
+      "loss": -0.0,
+      "num_tokens": 6812720.0,
+      "reward": -0.14257444441318512,
+      "reward_std": 0.29668545722961426,
+      "rewards/cosine_scaled_reward/mean": -0.14257442951202393,
+      "rewards/cosine_scaled_reward/std": 0.4257228672504425,
+      "step": 54
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1449.0,
+      "completions/mean_length": 1754.640625,
+      "completions/mean_terminated_length": 874.5625,
+      "completions/min_length": 581.0,
+      "completions/min_terminated_length": 581.0,
+      "epoch": 0.06285714285714286,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23320119082927704,
+      "learning_rate": 9.998245517681593e-07,
+      "loss": -0.0,
+      "num_tokens": 6935305.0,
+      "reward": -0.14078931510448456,
+      "reward_std": 0.17466726899147034,
+      "rewards/cosine_scaled_reward/mean": -0.14078931510448456,
+      "rewards/cosine_scaled_reward/std": 0.3331747353076935,
+      "step": 55
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.828125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1918.0,
+      "completions/mean_length": 1853.78125,
+      "completions/mean_terminated_length": 918.0,
+      "completions/min_length": 571.0,
+      "completions/min_terminated_length": 571.0,
+      "epoch": 0.064,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23405365645885468,
+      "learning_rate": 9.997258721585931e-07,
+      "loss": -0.0,
+      "num_tokens": 7064907.0,
+      "reward": -0.11611534655094147,
+      "reward_std": 0.19285616278648376,
+      "rewards/cosine_scaled_reward/mean": -0.11611534655094147,
+      "rewards/cosine_scaled_reward/std": 0.47406119108200073,
+      "step": 56
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1851.0,
+      "completions/mean_length": 1971.640625,
+      "completions/mean_terminated_length": 1437.125,
+      "completions/min_length": 1009.0,
+      "completions/min_terminated_length": 1009.0,
+      "epoch": 0.06514285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.20449356734752655,
+      "learning_rate": 9.996052735444862e-07,
+      "loss": 0.0,
+      "num_tokens": 7202660.0,
+      "reward": -0.27627938985824585,
+      "reward_std": 0.2080146074295044,
+      "rewards/cosine_scaled_reward/mean": -0.27627938985824585,
+      "rewards/cosine_scaled_reward/std": 0.2397139072418213,
+      "step": 57
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.65625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1813.0,
+      "completions/mean_length": 1678.09375,
+      "completions/mean_terminated_length": 971.9091186523438,
+      "completions/min_length": 540.0,
+      "completions/min_terminated_length": 540.0,
+      "epoch": 0.06628571428571428,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.251164048910141,
+      "learning_rate": 9.994627618036452e-07,
+      "loss": -0.0,
+      "num_tokens": 7320154.0,
+      "reward": -0.1333095282316208,
+      "reward_std": 0.27265745401382446,
+      "rewards/cosine_scaled_reward/mean": -0.1333095282316208,
+      "rewards/cosine_scaled_reward/std": 0.3821713328361511,
+      "step": 58
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1453.0,
+      "completions/mean_length": 1732.171875,
+      "completions/mean_terminated_length": 859.0,
+      "completions/min_length": 531.0,
+      "completions/min_terminated_length": 531.0,
+      "epoch": 0.06742857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22478283941745758,
+      "learning_rate": 9.992983438818915e-07,
+      "loss": -0.0,
+      "num_tokens": 7441477.0,
+      "reward": -0.18278491497039795,
+      "reward_std": 0.2154037207365036,
+      "rewards/cosine_scaled_reward/mean": -0.18278491497039795,
+      "rewards/cosine_scaled_reward/std": 0.3414745628833771,
+      "step": 59
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.765625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1935.0,
+      "completions/mean_length": 1798.375,
+      "completions/mean_terminated_length": 982.9334106445312,
+      "completions/min_length": 613.0,
+      "completions/min_terminated_length": 613.0,
+      "epoch": 0.06857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22602440416812897,
+      "learning_rate": 9.991120277927223e-07,
+      "loss": -0.0,
+      "num_tokens": 7567461.0,
+      "reward": -0.265900194644928,
+      "reward_std": 0.1530904918909073,
+      "rewards/cosine_scaled_reward/mean": -0.265900194644928,
+      "rewards/cosine_scaled_reward/std": 0.18254056572914124,
+      "step": 60
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.84375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1974.0,
+      "completions/mean_length": 1950.578125,
+      "completions/mean_terminated_length": 1424.5,
+      "completions/min_length": 808.0,
+      "completions/min_terminated_length": 808.0,
+      "epoch": 0.06971428571428571,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22684067487716675,
+      "learning_rate": 9.989038226169207e-07,
+      "loss": -0.0,
+      "num_tokens": 7703818.0,
+      "reward": -0.05269922316074371,
+      "reward_std": 0.3038993775844574,
+      "rewards/cosine_scaled_reward/mean": -0.052699219435453415,
+      "rewards/cosine_scaled_reward/std": 0.36445698142051697,
+      "step": 61
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.671875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1987.0,
+      "completions/mean_length": 1717.703125,
+      "completions/mean_terminated_length": 1041.3809814453125,
+      "completions/min_length": 432.0,
+      "completions/min_terminated_length": 432.0,
+      "epoch": 0.07085714285714285,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.23552638292312622,
+      "learning_rate": 9.98673738502114e-07,
+      "loss": 0.0,
+      "num_tokens": 7823983.0,
+      "reward": -0.07779724895954132,
+      "reward_std": 0.2913648784160614,
+      "rewards/cosine_scaled_reward/mean": -0.07779725641012192,
+      "rewards/cosine_scaled_reward/std": 0.4099881649017334,
+      "step": 62
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.484375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1751.0,
+      "completions/mean_length": 1600.625,
+      "completions/mean_terminated_length": 1180.3636474609375,
+      "completions/min_length": 420.0,
+      "completions/min_terminated_length": 420.0,
+      "epoch": 0.072,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.28230276703834534,
+      "learning_rate": 9.98421786662277e-07,
+      "loss": 0.0,
+      "num_tokens": 7936679.0,
+      "reward": -0.02632874622941017,
+      "reward_std": 0.25066205859184265,
+      "rewards/cosine_scaled_reward/mean": -0.02632874995470047,
+      "rewards/cosine_scaled_reward/std": 0.4263686537742615,
+      "step": 63
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.6875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2047.0,
+      "completions/mean_length": 1743.578125,
+      "completions/mean_terminated_length": 1073.8499755859375,
+      "completions/min_length": 496.0,
+      "completions/min_terminated_length": 496.0,
+      "epoch": 0.07314285714285715,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.266590416431427,
+      "learning_rate": 9.981479793771866e-07,
+      "loss": 0.0,
+      "num_tokens": 8059220.0,
+      "reward": -0.10920079052448273,
+      "reward_std": 0.3089619576931,
+      "rewards/cosine_scaled_reward/mean": -0.10920079052448273,
+      "rewards/cosine_scaled_reward/std": 0.43342384696006775,
+      "step": 64
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1127.0,
+      "completions/mean_length": 1690.609375,
+      "completions/mean_terminated_length": 618.4375,
+      "completions/min_length": 331.0,
+      "completions/min_terminated_length": 331.0,
+      "epoch": 0.07428571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2891872525215149,
+      "learning_rate": 9.97852329991824e-07,
+      "loss": 0.0,
+      "num_tokens": 8178123.0,
+      "reward": -0.2091352641582489,
+      "reward_std": 0.18792679905891418,
+      "rewards/cosine_scaled_reward/mean": -0.2091352641582489,
+      "rewards/cosine_scaled_reward/std": 0.40636762976646423,
+      "step": 65
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.46875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1529.0,
+      "completions/mean_length": 1320.453125,
+      "completions/mean_terminated_length": 678.5,
+      "completions/min_length": 219.0,
+      "completions/min_terminated_length": 219.0,
+      "epoch": 0.07542857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.30139341950416565,
+      "learning_rate": 9.975348529157229e-07,
+      "loss": -0.0,
+      "num_tokens": 8272864.0,
+      "reward": -0.012375831604003906,
+      "reward_std": 0.2539718747138977,
+      "rewards/cosine_scaled_reward/mean": -0.01237582415342331,
+      "rewards/cosine_scaled_reward/std": 0.45652061700820923,
+      "step": 66
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.953125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1368.0,
+      "completions/mean_length": 2001.21875,
+      "completions/mean_terminated_length": 1050.0,
+      "completions/min_length": 817.0,
+      "completions/min_terminated_length": 817.0,
+      "epoch": 0.07657142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21435414254665375,
+      "learning_rate": 9.971955636222684e-07,
+      "loss": 0.0,
+      "num_tokens": 8411678.0,
+      "reward": -0.27966073155403137,
+      "reward_std": 0.14496129751205444,
+      "rewards/cosine_scaled_reward/mean": -0.27966073155403137,
+      "rewards/cosine_scaled_reward/std": 0.1733873188495636,
+      "step": 67
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.453125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1963.0,
+      "completions/mean_length": 1359.65625,
+      "completions/mean_terminated_length": 789.3142700195312,
+      "completions/min_length": 347.0,
+      "completions/min_terminated_length": 347.0,
+      "epoch": 0.07771428571428571,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.3244759440422058,
+      "learning_rate": 9.968344786479415e-07,
+      "loss": -0.0,
+      "num_tokens": 8507952.0,
+      "reward": -0.06231251358985901,
+      "reward_std": 0.31347835063934326,
+      "rewards/cosine_scaled_reward/mean": -0.062312521040439606,
+      "rewards/cosine_scaled_reward/std": 0.40184450149536133,
+      "step": 68
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.609375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1859.0,
+      "completions/mean_length": 1572.78125,
+      "completions/mean_terminated_length": 831.4400024414062,
+      "completions/min_length": 358.0,
+      "completions/min_terminated_length": 358.0,
+      "epoch": 0.07885714285714286,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.3429071605205536,
+      "learning_rate": 9.964516155915151e-07,
+      "loss": 0.0,
+      "num_tokens": 8618954.0,
+      "reward": -0.24097035825252533,
+      "reward_std": 0.22784993052482605,
+      "rewards/cosine_scaled_reward/mean": -0.24097035825252533,
+      "rewards/cosine_scaled_reward/std": 0.2594495415687561,
+      "step": 69
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.828125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1921.0,
+      "completions/mean_length": 1859.578125,
+      "completions/mean_terminated_length": 951.727294921875,
+      "completions/min_length": 532.0,
+      "completions/min_terminated_length": 532.0,
+      "epoch": 0.08,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.221941277384758,
+      "learning_rate": 9.960469931131936e-07,
+      "loss": -0.0,
+      "num_tokens": 8749423.0,
+      "reward": -0.27105003595352173,
+      "reward_std": 0.16835230588912964,
+      "rewards/cosine_scaled_reward/mean": -0.27105003595352173,
+      "rewards/cosine_scaled_reward/std": 0.21196867525577545,
+      "step": 70
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.6875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1629.0,
+      "completions/mean_length": 1668.265625,
+      "completions/mean_terminated_length": 832.8500366210938,
+      "completions/min_length": 489.0,
+      "completions/min_terminated_length": 489.0,
+      "epoch": 0.08114285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2909034192562103,
+      "learning_rate": 9.956206309337066e-07,
+      "loss": -0.0,
+      "num_tokens": 8866912.0,
+      "reward": -0.09497882425785065,
+      "reward_std": 0.2813299000263214,
+      "rewards/cosine_scaled_reward/mean": -0.09497880935668945,
+      "rewards/cosine_scaled_reward/std": 0.4832696318626404,
+      "step": 71
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.6875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1903.0,
+      "completions/mean_length": 1697.671875,
+      "completions/mean_terminated_length": 926.9500122070312,
+      "completions/min_length": 275.0,
+      "completions/min_terminated_length": 275.0,
+      "epoch": 0.08228571428571428,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.3016415238380432,
+      "learning_rate": 9.951725498333448e-07,
+      "loss": -0.0,
+      "num_tokens": 8985915.0,
+      "reward": -0.22967606782913208,
+      "reward_std": 0.18875859677791595,
+      "rewards/cosine_scaled_reward/mean": -0.2296760529279709,
+      "rewards/cosine_scaled_reward/std": 0.22012120485305786,
+      "step": 72
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.953125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1957.0,
+      "completions/mean_length": 2020.703125,
+      "completions/mean_terminated_length": 1465.666748046875,
+      "completions/min_length": 1143.0,
+      "completions/min_terminated_length": 1143.0,
+      "epoch": 0.08342857142857144,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21586637198925018,
+      "learning_rate": 9.947027716509488e-07,
+      "loss": 0.0,
+      "num_tokens": 9125968.0,
+      "reward": -0.24284613132476807,
+      "reward_std": 0.22862236201763153,
+      "rewards/cosine_scaled_reward/mean": -0.24284613132476807,
+      "rewards/cosine_scaled_reward/std": 0.24740919470787048,
+      "step": 73
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.890625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1955.0,
+      "completions/mean_length": 1975.09375,
+      "completions/mean_terminated_length": 1381.4285888671875,
+      "completions/min_length": 532.0,
+      "completions/min_terminated_length": 532.0,
+      "epoch": 0.08457142857142858,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.21680164337158203,
+      "learning_rate": 9.942113192828444e-07,
+      "loss": 0.0,
+      "num_tokens": 9262302.0,
+      "reward": -0.1543380469083786,
+      "reward_std": 0.24083258211612701,
+      "rewards/cosine_scaled_reward/mean": -0.1543380618095398,
+      "rewards/cosine_scaled_reward/std": 0.3356986939907074,
+      "step": 74
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.71875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2038.0,
+      "completions/mean_length": 1836.3125,
+      "completions/mean_terminated_length": 1295.3333740234375,
+      "completions/min_length": 653.0,
+      "completions/min_terminated_length": 653.0,
+      "epoch": 0.08571428571428572,
+      "frac_reward_zero_std": 0.125,
+      "grad_norm": 0.1845395565032959,
+      "learning_rate": 9.93698216681727e-07,
+      "loss": 0.0,
+      "num_tokens": 9390786.0,
+      "reward": -0.12792138755321503,
+      "reward_std": 0.10224759578704834,
+      "rewards/cosine_scaled_reward/mean": -0.12792138755321503,
+      "rewards/cosine_scaled_reward/std": 0.4530969560146332,
+      "step": 75
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.765625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1422.0,
+      "completions/mean_length": 1764.109375,
+      "completions/mean_terminated_length": 836.7333984375,
+      "completions/min_length": 320.0,
+      "completions/min_terminated_length": 320.0,
+      "epoch": 0.08685714285714285,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.26535236835479736,
+      "learning_rate": 9.931634888554935e-07,
+      "loss": -0.0,
+      "num_tokens": 9514089.0,
+      "reward": -0.27717918157577515,
+      "reward_std": 0.19932743906974792,
+      "rewards/cosine_scaled_reward/mean": -0.27717918157577515,
+      "rewards/cosine_scaled_reward/std": 0.20844916999340057,
+      "step": 76
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1864.0,
+      "completions/mean_length": 1945.109375,
+      "completions/mean_terminated_length": 1224.875,
+      "completions/min_length": 702.0,
+      "completions/min_terminated_length": 702.0,
+      "epoch": 0.088,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2047174870967865,
+      "learning_rate": 9.926071618660237e-07,
+      "loss": -0.0,
+      "num_tokens": 9650152.0,
+      "reward": -0.09873012453317642,
+      "reward_std": 0.22244854271411896,
+      "rewards/cosine_scaled_reward/mean": -0.09873010218143463,
+      "rewards/cosine_scaled_reward/std": 0.34491515159606934,
+      "step": 77
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.859375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1976.0,
+      "completions/mean_length": 1928.703125,
+      "completions/mean_terminated_length": 1199.6666259765625,
+      "completions/min_length": 722.0,
+      "completions/min_terminated_length": 722.0,
+      "epoch": 0.08914285714285715,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22559019923210144,
+      "learning_rate": 9.9202926282791e-07,
+      "loss": 0.0,
+      "num_tokens": 9784309.0,
+      "reward": -0.09572747349739075,
+      "reward_std": 0.23068635165691376,
+      "rewards/cosine_scaled_reward/mean": -0.09572747349739075,
+      "rewards/cosine_scaled_reward/std": 0.38660773634910583,
+      "step": 78
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.5625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1864.0,
+      "completions/mean_length": 1508.40625,
+      "completions/mean_terminated_length": 814.6428833007812,
+      "completions/min_length": 339.0,
+      "completions/min_terminated_length": 339.0,
+      "epoch": 0.09028571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24668477475643158,
+      "learning_rate": 9.91429819907136e-07,
+      "loss": 0.0,
+      "num_tokens": 9890943.0,
+      "reward": -0.1618795394897461,
+      "reward_std": 0.22540031373500824,
+      "rewards/cosine_scaled_reward/mean": -0.1618795245885849,
+      "rewards/cosine_scaled_reward/std": 0.3233039081096649,
+      "step": 79
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.890625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1996.0,
+      "completions/mean_length": 2012.671875,
+      "completions/mean_terminated_length": 1725.0001220703125,
+      "completions/min_length": 1283.0,
+      "completions/min_terminated_length": 1283.0,
+      "epoch": 0.09142857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24282054603099823,
+      "learning_rate": 9.908088623197048e-07,
+      "loss": -0.0,
+      "num_tokens": 10030146.0,
+      "reward": -0.25591158866882324,
+      "reward_std": 0.15104801952838898,
+      "rewards/cosine_scaled_reward/mean": -0.25591158866882324,
+      "rewards/cosine_scaled_reward/std": 0.18741995096206665,
+      "step": 80
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.796875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1878.0,
+      "completions/mean_length": 1821.921875,
+      "completions/mean_terminated_length": 935.0000610351562,
+      "completions/min_length": 580.0,
+      "completions/min_terminated_length": 580.0,
+      "epoch": 0.09257142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.3027254641056061,
+      "learning_rate": 9.901664203302124e-07,
+      "loss": 0.0,
+      "num_tokens": 10158021.0,
+      "reward": -0.15331333875656128,
+      "reward_std": 0.18424856662750244,
+      "rewards/cosine_scaled_reward/mean": -0.15331333875656128,
+      "rewards/cosine_scaled_reward/std": 0.24023762345314026,
+      "step": 81
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.703125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1852.0,
+      "completions/mean_length": 1734.28125,
+      "completions/mean_terminated_length": 991.26318359375,
+      "completions/min_length": 477.0,
+      "completions/min_terminated_length": 477.0,
+      "epoch": 0.09371428571428571,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2466808557510376,
+      "learning_rate": 9.895025252503755e-07,
+      "loss": -0.0,
+      "num_tokens": 10279343.0,
+      "reward": -0.07192108780145645,
+      "reward_std": 0.2587333917617798,
+      "rewards/cosine_scaled_reward/mean": -0.07192108780145645,
+      "rewards/cosine_scaled_reward/std": 0.46087121963500977,
+      "step": 82
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.671875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1571.0,
+      "completions/mean_length": 1688.78125,
+      "completions/mean_terminated_length": 953.2380981445312,
+      "completions/min_length": 495.0,
+      "completions/min_terminated_length": 495.0,
+      "epoch": 0.09485714285714286,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2600877285003662,
+      "learning_rate": 9.888172094375033e-07,
+      "loss": 0.0,
+      "num_tokens": 10398513.0,
+      "reward": -0.1718086451292038,
+      "reward_std": 0.2223512828350067,
+      "rewards/cosine_scaled_reward/mean": -0.1718086451292038,
+      "rewards/cosine_scaled_reward/std": 0.2828122675418854,
+      "step": 83
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.84375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1666.0,
+      "completions/mean_length": 1838.203125,
+      "completions/mean_terminated_length": 705.2999877929688,
+      "completions/min_length": 435.0,
+      "completions/min_terminated_length": 435.0,
+      "epoch": 0.096,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.22531215846538544,
+      "learning_rate": 9.881105062929221e-07,
+      "loss": 0.0,
+      "num_tokens": 10526854.0,
+      "reward": -0.2154863476753235,
+      "reward_std": 0.261901319026947,
+      "rewards/cosine_scaled_reward/mean": -0.2154863476753235,
+      "rewards/cosine_scaled_reward/std": 0.29268571734428406,
+      "step": 84
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.8125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2026.0,
+      "completions/mean_length": 1926.34375,
+      "completions/mean_terminated_length": 1399.166748046875,
+      "completions/min_length": 880.0,
+      "completions/min_terminated_length": 880.0,
+      "epoch": 0.09714285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.19031891226768494,
+      "learning_rate": 9.873824502603459e-07,
+      "loss": -0.0,
+      "num_tokens": 10660460.0,
+      "reward": -0.21009978652000427,
+      "reward_std": 0.19575349986553192,
+      "rewards/cosine_scaled_reward/mean": -0.21009978652000427,
+      "rewards/cosine_scaled_reward/std": 0.2456056773662567,
+      "step": 85
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.734375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1920.0,
+      "completions/mean_length": 1713.21875,
+      "completions/mean_terminated_length": 787.6470336914062,
+      "completions/min_length": 547.0,
+      "completions/min_terminated_length": 547.0,
+      "epoch": 0.09828571428571428,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.258359432220459,
+      "learning_rate": 9.866330768241983e-07,
+      "loss": -0.0,
+      "num_tokens": 10780962.0,
+      "reward": -0.1955144852399826,
+      "reward_std": 0.24323132634162903,
+      "rewards/cosine_scaled_reward/mean": -0.1955144852399826,
+      "rewards/cosine_scaled_reward/std": 0.3071554899215698,
+      "step": 86
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1815.0,
+      "completions/mean_length": 1656.0,
+      "completions/mean_terminated_length": 1002.6666870117188,
+      "completions/min_length": 532.0,
+      "completions/min_terminated_length": 532.0,
+      "epoch": 0.09942857142857142,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2636864185333252,
+      "learning_rate": 9.85862422507884e-07,
+      "loss": 0.0,
+      "num_tokens": 10897066.0,
+      "reward": -0.1988150179386139,
+      "reward_std": 0.24088150262832642,
+      "rewards/cosine_scaled_reward/mean": -0.1988150179386139,
+      "rewards/cosine_scaled_reward/std": 0.2925129532814026,
+      "step": 87
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.703125,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1802.0,
+      "completions/mean_length": 1755.234375,
+      "completions/mean_terminated_length": 1061.8421630859375,
+      "completions/min_length": 500.0,
+      "completions/min_terminated_length": 500.0,
+      "epoch": 0.10057142857142858,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.29112017154693604,
+      "learning_rate": 9.850705248720068e-07,
+      "loss": 0.0,
+      "num_tokens": 11019913.0,
+      "reward": -0.02967459335923195,
+      "reward_std": 0.3240855932235718,
+      "rewards/cosine_scaled_reward/mean": -0.029674597084522247,
+      "rewards/cosine_scaled_reward/std": 0.3718070983886719,
+      "step": 88
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.671875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1768.0,
+      "completions/mean_length": 1752.78125,
+      "completions/mean_terminated_length": 1148.2857666015625,
+      "completions/min_length": 619.0,
+      "completions/min_terminated_length": 619.0,
+      "epoch": 0.10171428571428572,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2357943207025528,
+      "learning_rate": 9.8425742251254e-07,
+      "loss": -0.0,
+      "num_tokens": 11143091.0,
+      "reward": -0.1188301220536232,
+      "reward_std": 0.296513170003891,
+      "rewards/cosine_scaled_reward/mean": -0.1188301220536232,
+      "rewards/cosine_scaled_reward/std": 0.3878798484802246,
+      "step": 89
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.5625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1693.0,
+      "completions/mean_length": 1633.84375,
+      "completions/mean_terminated_length": 1101.357177734375,
+      "completions/min_length": 568.0,
+      "completions/min_terminated_length": 568.0,
+      "epoch": 0.10285714285714286,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.32384219765663147,
+      "learning_rate": 9.83423155058946e-07,
+      "loss": -0.0,
+      "num_tokens": 11257657.0,
+      "reward": -0.22837099432945251,
+      "reward_std": 0.18625205755233765,
+      "rewards/cosine_scaled_reward/mean": -0.22837099432945251,
+      "rewards/cosine_scaled_reward/std": 0.23636196553707123,
+      "step": 90
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1833.0,
+      "completions/mean_length": 1847.21875,
+      "completions/mean_terminated_length": 1244.875,
+      "completions/min_length": 716.0,
+      "completions/min_terminated_length": 716.0,
+      "epoch": 0.104,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.24563109874725342,
+      "learning_rate": 9.825677631722435e-07,
+      "loss": -0.0,
+      "num_tokens": 11386447.0,
+      "reward": -0.11780542880296707,
+      "reward_std": 0.3100074827671051,
+      "rewards/cosine_scaled_reward/mean": -0.11780542135238647,
+      "rewards/cosine_scaled_reward/std": 0.39149248600006104,
+      "step": 91
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.609375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1462.0,
+      "completions/mean_length": 1595.125,
+      "completions/mean_terminated_length": 888.6399536132812,
+      "completions/min_length": 464.0,
+      "completions/min_terminated_length": 464.0,
+      "epoch": 0.10514285714285715,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2472057044506073,
+      "learning_rate": 9.816912885430258e-07,
+      "loss": -0.0,
+      "num_tokens": 11498527.0,
+      "reward": -0.2128506749868393,
+      "reward_std": 0.20926561951637268,
+      "rewards/cosine_scaled_reward/mean": -0.2128506898880005,
+      "rewards/cosine_scaled_reward/std": 0.23348061740398407,
+      "step": 92
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.9375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1221.0,
+      "completions/mean_length": 1979.953125,
+      "completions/mean_terminated_length": 959.25,
+      "completions/min_length": 822.0,
+      "completions/min_terminated_length": 822.0,
+      "epoch": 0.10628571428571429,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2550150454044342,
+      "learning_rate": 9.807937738894303e-07,
+      "loss": -0.0,
+      "num_tokens": 11636588.0,
+      "reward": -0.2922024428844452,
+      "reward_std": 0.1515069603919983,
+      "rewards/cosine_scaled_reward/mean": -0.2922024726867676,
+      "rewards/cosine_scaled_reward/std": 0.18899379670619965,
+      "step": 93
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.796875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1544.0,
+      "completions/mean_length": 1830.609375,
+      "completions/mean_terminated_length": 977.769287109375,
+      "completions/min_length": 533.0,
+      "completions/min_terminated_length": 533.0,
+      "epoch": 0.10742857142857143,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.27164825797080994,
+      "learning_rate": 9.798752629550546e-07,
+      "loss": -0.0,
+      "num_tokens": 11763515.0,
+      "reward": -0.18001651763916016,
+      "reward_std": 0.18973413109779358,
+      "rewards/cosine_scaled_reward/mean": -0.18001650273799896,
+      "rewards/cosine_scaled_reward/std": 0.4316568076610565,
+      "step": 94
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.921875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2034.0,
+      "completions/mean_length": 2004.671875,
+      "completions/mean_terminated_length": 1493.4000244140625,
+      "completions/min_length": 960.0,
+      "completions/min_terminated_length": 960.0,
+      "epoch": 0.10857142857142857,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.20861269533634186,
+      "learning_rate": 9.78935800506826e-07,
+      "loss": -0.0,
+      "num_tokens": 11902342.0,
+      "reward": -0.24148261547088623,
+      "reward_std": 0.18629083037376404,
+      "rewards/cosine_scaled_reward/mean": -0.24148263037204742,
+      "rewards/cosine_scaled_reward/std": 0.23122739791870117,
+      "step": 95
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.6875,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1298.0,
+      "completions/mean_length": 1703.359375,
+      "completions/mean_terminated_length": 945.1500244140625,
+      "completions/min_length": 455.0,
+      "completions/min_terminated_length": 455.0,
+      "epoch": 0.10971428571428571,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2585296928882599,
+      "learning_rate": 9.779754323328192e-07,
+      "loss": 0.0,
+      "num_tokens": 12022493.0,
+      "reward": -0.11465626955032349,
+      "reward_std": 0.24939197301864624,
+      "rewards/cosine_scaled_reward/mean": -0.11465626955032349,
+      "rewards/cosine_scaled_reward/std": 0.4384477138519287,
+      "step": 96
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.75,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1809.0,
+      "completions/mean_length": 1819.921875,
+      "completions/mean_terminated_length": 1135.6875,
+      "completions/min_length": 425.0,
+      "completions/min_terminated_length": 425.0,
+      "epoch": 0.11085714285714286,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.3019813299179077,
+      "learning_rate": 9.769942052400235e-07,
+      "loss": 0.0,
+      "num_tokens": 12149232.0,
+      "reward": -0.18846748769283295,
+      "reward_std": 0.2666187584400177,
+      "rewards/cosine_scaled_reward/mean": -0.18846750259399414,
+      "rewards/cosine_scaled_reward/std": 0.3043021559715271,
+      "step": 97
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.609375,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 2003.0,
+      "completions/mean_length": 1677.296875,
+      "completions/mean_terminated_length": 1099.0,
+      "completions/min_length": 315.0,
+      "completions/min_terminated_length": 315.0,
+      "epoch": 0.112,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2722402513027191,
+      "learning_rate": 9.759921670520634e-07,
+      "loss": 0.0,
+      "num_tokens": 12267643.0,
+      "reward": -0.09557384252548218,
+      "reward_std": 0.2643275558948517,
+      "rewards/cosine_scaled_reward/mean": -0.09557383507490158,
+      "rewards/cosine_scaled_reward/std": 0.3361329138278961,
+      "step": 98
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.765625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1141.0,
+      "completions/mean_length": 1716.59375,
+      "completions/mean_terminated_length": 634.0000610351562,
+      "completions/min_length": 393.0,
+      "completions/min_terminated_length": 393.0,
+      "epoch": 0.11314285714285714,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2789485454559326,
+      "learning_rate": 9.749693666068663e-07,
+      "loss": -0.0,
+      "num_tokens": 12388673.0,
+      "reward": -0.11132554709911346,
+      "reward_std": 0.1736970841884613,
+      "rewards/cosine_scaled_reward/mean": -0.11132554709911346,
+      "rewards/cosine_scaled_reward/std": 0.38663193583488464,
+      "step": 99
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.625,
+      "completions/max_length": 2048.0,
+      "completions/max_terminated_length": 1906.0,
+      "completions/mean_length": 1627.78125,
+      "completions/mean_terminated_length": 927.4166870117188,
+      "completions/min_length": 426.0,
+      "completions/min_terminated_length": 426.0,
+      "epoch": 0.11428571428571428,
+      "frac_reward_zero_std": 0.0,
+      "grad_norm": 0.2479974329471588,
+      "learning_rate": 9.739258537542835e-07,
+      "loss": 0.0,
+      "num_tokens": 12502563.0,
+      "reward": 0.05247430503368378,
+      "reward_std": 0.2633323669433594,
+      "rewards/cosine_scaled_reward/mean": 0.05247429758310318,
+      "rewards/cosine_scaled_reward/std": 0.44700634479522705,
+      "step": 100
+    }
+  ],
+  "logging_steps": 1,
+  "max_steps": 500,
+  "num_input_tokens_seen": 12502563,
+  "num_train_epochs": 1,
+  "save_steps": 50,
+  "stateful_callbacks": {
+    "TrainerControl": {
+      "args": {
+        "should_epoch_stop": false,
+        "should_evaluate": false,
+        "should_log": false,
+        "should_save": true,
+        "should_training_stop": false
+      },
+      "attributes": {}
+    }
+  },
+  "total_flos": 0.0,
+  "train_batch_size": 4,
+  "trial_name": null,
+  "trial_params": null
+}

checkpoint-100/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ef86752cf6cf52e796d735b4903ab569bc33d2865692788b12aadf084a12139e
+size 8824

checkpoint-100/zero_to_fp32.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+
+# Copyright (c) Microsoft Corporation.
+# SPDX-License-Identifier: Apache-2.0
+
+# DeepSpeed Team
+
+# This script extracts fp32 consolidated weights from a zero 1, 2 and 3 DeepSpeed checkpoints. It gets
+# copied into the top level checkpoint dir, so the user can easily do the conversion at any point in
+# the future. Once extracted, the weights don't require DeepSpeed and can be used in any
+# application.
+#
+# example:
+#   python zero_to_fp32.py . output_dir/
+#   or
+#   python zero_to_fp32.py . output_dir/ --safe_serialization
+
+import argparse
+import torch
+import glob
+import math
+import os
+import re
+import gc
+import json
+import numpy as np
+from tqdm import tqdm
+from collections import OrderedDict
+from dataclasses import dataclass
+
+# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
+# DeepSpeed data structures it has to be available in the current python environment.
+from deepspeed.utils import logger
+from deepspeed.checkpoint.constants import (DS_VERSION, OPTIMIZER_STATE_DICT, SINGLE_PARTITION_OF_FP32_GROUPS,
+                                            FP32_FLAT_GROUPS, ZERO_STAGE, PARTITION_COUNT, PARAM_SHAPES, BUFFER_NAMES,
+                                            FROZEN_PARAM_SHAPES, FROZEN_PARAM_FRAGMENTS)
+
+
+@dataclass
+class zero_model_state:
+    buffers: dict()
+    param_shapes: dict()
+    shared_params: list
+    ds_version: int
+    frozen_param_shapes: dict()
+    frozen_param_fragments: dict()
+
+
+debug = 0
+
+# load to cpu
+device = torch.device('cpu')
+
+
+def atoi(text):
+    return int(text) if text.isdigit() else text
+
+
+def natural_keys(text):
+    '''
+    alist.sort(key=natural_keys) sorts in human order
+    http://nedbatchelder.com/blog/200712/human_sorting.html
+    (See Toothy's implementation in the comments)
+    '''
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
+
+
+def get_model_state_file(checkpoint_dir, zero_stage):
+    if not os.path.isdir(checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{checkpoint_dir}' doesn't exist")
+
+    # there should be only one file
+    if zero_stage <= 2:
+        file = os.path.join(checkpoint_dir, "mp_rank_00_model_states.pt")
+    elif zero_stage == 3:
+        file = os.path.join(checkpoint_dir, "zero_pp_rank_0_mp_rank_00_model_states.pt")
+
+    if not os.path.exists(file):
+        raise FileNotFoundError(f"can't find model states file at '{file}'")
+
+    return file
+
+
+def get_checkpoint_files(checkpoint_dir, glob_pattern):
+    # XXX: need to test that this simple glob rule works for multi-node setup too
+    ckpt_files = sorted(glob.glob(os.path.join(checkpoint_dir, glob_pattern)), key=natural_keys)
+
+    if len(ckpt_files) == 0:
+        raise FileNotFoundError(f"can't find {glob_pattern} files in directory '{checkpoint_dir}'")
+
+    return ckpt_files
+
+
+def get_optim_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_optim_states.pt")
+
+
+def get_model_state_files(checkpoint_dir):
+    return get_checkpoint_files(checkpoint_dir, "*_model_states.pt")
+
+
+def parse_model_states(files):
+    zero_model_states = []
+    for file in files:
+        state_dict = torch.load(file, map_location=device, weights_only=False)
+
+        if BUFFER_NAMES not in state_dict:
+            raise ValueError(f"{file} is not a model state checkpoint")
+        buffer_names = state_dict[BUFFER_NAMES]
+        if debug:
+            print("Found buffers:", buffer_names)
+
+        # recover just the buffers while restoring them to fp32 if they were saved in fp16
+        buffers = {k: v.float() for k, v in state_dict["module"].items() if k in buffer_names}
+        param_shapes = state_dict[PARAM_SHAPES]
+
+        # collect parameters that are included in param_shapes
+        param_names = []
+        for s in param_shapes:
+            for name in s.keys():
+                param_names.append(name)
+
+        # update with frozen parameters
+        frozen_param_shapes = state_dict.get(FROZEN_PARAM_SHAPES, None)
+        if frozen_param_shapes is not None:
+            if debug:
+                print(f"Found frozen_param_shapes: {frozen_param_shapes}")
+            param_names += list(frozen_param_shapes.keys())
+
+        # handle shared params
+        shared_params = [[k, v] for k, v in state_dict["shared_params"].items()]
+
+        ds_version = state_dict.get(DS_VERSION, None)
+
+        frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
+
+        z_model_state = zero_model_state(buffers=buffers,
+                                         param_shapes=param_shapes,
+                                         shared_params=shared_params,
+                                         ds_version=ds_version,
+                                         frozen_param_shapes=frozen_param_shapes,
+                                         frozen_param_fragments=frozen_param_fragments)
+        zero_model_states.append(z_model_state)
+
+    return zero_model_states
+
+
+def parse_optim_states(files, ds_checkpoint_dir):
+    total_files = len(files)
+    state_dicts = []
+    for f in tqdm(files, desc='Loading checkpoint shards'):
+        state_dict = torch.load(f, map_location=device, mmap=True, weights_only=False)
+        # immediately discard the potentially huge 2 optimizer states as we only care for fp32 master weights
+        # and also handle the case where it was already removed by another helper script
+        state_dict["optimizer_state_dict"].pop("optimizer_state_dict", None)
+        state_dicts.append(state_dict)
+
+    if not ZERO_STAGE in state_dicts[0][OPTIMIZER_STATE_DICT]:
+        raise ValueError(f"{files[0]} is not a zero checkpoint")
+    zero_stage = state_dicts[0][OPTIMIZER_STATE_DICT][ZERO_STAGE]
+    world_size = state_dicts[0][OPTIMIZER_STATE_DICT][PARTITION_COUNT]
+
+    # For ZeRO-2 each param group can have different partition_count as data parallelism for expert
+    # parameters can be different from data parallelism for non-expert parameters. So we can just
+    # use the max of the partition_count to get the dp world_size.
+
+    if type(world_size) is list:
+        world_size = max(world_size)
+
+    if world_size != total_files:
+        raise ValueError(
+            f"Expected {world_size} of '*_optim_states.pt' under '{ds_checkpoint_dir}' but found {total_files} files. "
+            "Possibly due to an overwrite of an old checkpoint, or a checkpoint didn't get saved by one or more processes."
+        )
+
+    # the groups are named differently in each stage
+    if zero_stage <= 2:
+        fp32_groups_key = SINGLE_PARTITION_OF_FP32_GROUPS
+    elif zero_stage == 3:
+        fp32_groups_key = FP32_FLAT_GROUPS
+    else:
+        raise ValueError(f"unknown zero stage {zero_stage}")
+
+    fp32_flat_groups = [state_dicts[i][OPTIMIZER_STATE_DICT][fp32_groups_key] for i in range(len(state_dicts))]
+    return zero_stage, world_size, fp32_flat_groups
+
+
+def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters):
+    """
+    Returns fp32 state_dict reconstructed from ds checkpoint
+
+    Args:
+        - ``ds_checkpoint_dir``: path to the deepspeed checkpoint folder (where the optimizer files are)
+
+    """
+    print(f"Processing zero checkpoint '{ds_checkpoint_dir}'")
+
+    optim_files = get_optim_files(ds_checkpoint_dir)
+    zero_stage, world_size, fp32_flat_groups = parse_optim_states(optim_files, ds_checkpoint_dir)
+    print(f"Detected checkpoint of type zero stage {zero_stage}, world_size: {world_size}")
+
+    model_files = get_model_state_files(ds_checkpoint_dir)
+
+    zero_model_states = parse_model_states(model_files)
+    print(f'Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}')
+
+    if zero_stage <= 2:
+        return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+    elif zero_stage == 3:
+        return _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                                          exclude_frozen_parameters)
+
+
+def _zero2_merge_frozen_params(state_dict, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+    frozen_param_fragments = zero_model_states[0].frozen_param_fragments
+
+    if debug:
+        num_elem = sum(s.numel() for s in frozen_param_shapes.values())
+        print(f'rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        state_dict[name] = frozen_param_fragments[name]
+
+        if debug:
+            print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _has_callable(obj, fn):
+    attr = getattr(obj, fn, None)
+    return callable(attr)
+
+
+def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+
+    # Reconstruction protocol:
+    #
+    # XXX: document this
+
+    if debug:
+        for i in range(world_size):
+            for j in range(len(fp32_flat_groups[0])):
+                print(f"{FP32_FLAT_GROUPS}[{i}][{j}].shape={fp32_flat_groups[i][j].shape}")
+
+    # XXX: memory usage doubles here (zero2)
+    num_param_groups = len(fp32_flat_groups[0])
+    merged_single_partition_of_fp32_groups = []
+    for i in range(num_param_groups):
+        merged_partitions = [sd[i] for sd in fp32_flat_groups]
+        full_single_fp32_vector = torch.cat(merged_partitions, 0)
+        merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
+    avail_numel = sum(
+        [full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups])
+
+    if debug:
+        wanted_params = sum([len(shapes) for shapes in param_shapes])
+        wanted_numel = sum([sum(shape.numel() for shape in shapes.values()) for shapes in param_shapes])
+        # not asserting if there is a mismatch due to possible padding
+        print(f"Have {avail_numel} numels to process.")
+        print(f"Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    total_numel = 0
+    total_params = 0
+    for shapes, full_single_fp32_vector in zip(param_shapes, merged_single_partition_of_fp32_groups):
+        offset = 0
+        avail_numel = full_single_fp32_vector.numel()
+        for name, shape in shapes.items():
+
+            unpartitioned_numel = shape.numel() if _has_callable(shape, 'numel') else math.prod(shape)
+            total_numel += unpartitioned_numel
+            total_params += 1
+
+            if debug:
+                print(f"{name} full shape: {shape} unpartitioned numel {unpartitioned_numel} ")
+            state_dict[name] = full_single_fp32_vector.narrow(0, offset, unpartitioned_numel).view(shape)
+            offset += unpartitioned_numel
+
+        # Z2 started to align to 2*world_size to improve nccl performance. Therefore both offset and
+        # avail_numel can differ by anywhere between 0..2*world_size. Due to two unrelated complex
+        # paddings performed in the code it's almost impossible to predict the exact numbers w/o the
+        # live optimizer object, so we are checking that the numbers are within the right range
+        align_to = 2 * world_size
+
+        def zero2_align(x):
+            return align_to * math.ceil(x / align_to)
+
+        if debug:
+            print(f"original offset={offset}, avail_numel={avail_numel}")
+
+        offset = zero2_align(offset)
+        avail_numel = zero2_align(avail_numel)
+
+        if debug:
+            print(f"aligned  offset={offset}, avail_numel={avail_numel}")
+
+        # Sanity check
+        if offset != avail_numel:
+            raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero2_merge_frozen_params(state_dict, zero_model_states)
+
+    _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def zero3_partitioned_param_info(unpartitioned_numel, world_size):
+    remainder = unpartitioned_numel % world_size
+    padding_numel = (world_size - remainder) if remainder else 0
+    partitioned_numel = math.ceil(unpartitioned_numel / world_size)
+    return partitioned_numel, padding_numel
+
+
+def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
+    if zero_model_states[0].frozen_param_shapes is None or len(zero_model_states[0].frozen_param_shapes) == 0:
+        return
+
+    if debug:
+        for i in range(world_size):
+            num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
+            print(f'rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}')
+
+        frozen_param_shapes = zero_model_states[0].frozen_param_shapes
+        wanted_params = len(frozen_param_shapes)
+        wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
+        avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
+        print(f'Frozen params: Have {avail_numel} numels to process.')
+        print(f'Frozen params: Need {wanted_numel} numels in {wanted_params} params')
+
+    total_params = 0
+    total_numel = 0
+    for name, shape in zero_model_states[0].frozen_param_shapes.items():
+        total_params += 1
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+
+        param_frags = tuple(model_state.frozen_param_fragments[name] for model_state in zero_model_states)
+        state_dict[name] = torch.cat(param_frags, 0).narrow(0, 0, unpartitioned_numel).view(shape)
+
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Frozen params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+    print(f"Reconstructed Frozen fp32 state dict with {total_params} params {total_numel} elements")
+
+
+class GatheredTensor:
+    """
+    A pseudo tensor that collects partitioned weights.
+    It is more memory efficient when there are multiple groups.
+    """
+
+    def __init__(self, flat_groups, flat_groups_offset, offset, partitioned_numel, shape):
+        self.flat_groups = flat_groups
+        self.flat_groups_offset = flat_groups_offset
+        self.offset = offset
+        self.partitioned_numel = partitioned_numel
+        self.shape = shape
+        self.dtype = self.flat_groups[0][0].dtype
+
+    def contiguous(self):
+        """
+        Merge partitioned weights from flat_groups into a single tensor.
+        """
+        end_idx = self.offset + self.partitioned_numel
+        world_size = len(self.flat_groups)
+        pad_flat_param_chunks = []
+
+        for rank_i in range(world_size):
+            # for each rank, we need to collect weights from related group/groups
+            flat_groups_at_rank_i = self.flat_groups[rank_i]
+            start_group_id = None
+            end_group_id = None
+            for group_id in range(len(self.flat_groups_offset)):
+                if self.flat_groups_offset[group_id] <= self.offset < self.flat_groups_offset[group_id + 1]:
+                    start_group_id = group_id
+                if self.flat_groups_offset[group_id] < end_idx <= self.flat_groups_offset[group_id + 1]:
+                    end_group_id = group_id
+                    break
+            # collect weights from related group/groups
+            for group_id in range(start_group_id, end_group_id + 1):
+                flat_tensor = flat_groups_at_rank_i[group_id]
+                start_offset = self.offset - self.flat_groups_offset[group_id]
+                end_offset = min(end_idx, self.flat_groups_offset[group_id + 1]) - self.flat_groups_offset[group_id]
+                pad_flat_param_chunks.append(flat_tensor[start_offset:end_offset])
+
+        # collect weights from all ranks
+        pad_flat_param = torch.cat(pad_flat_param_chunks, dim=0)
+        param = pad_flat_param[:self.shape.numel()].view(self.shape).contiguous()
+        return param
+
+
+def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states):
+    param_shapes = zero_model_states[0].param_shapes
+    avail_numel = sum([flat_group.numel() for flat_group in fp32_flat_groups[0]]) * world_size
+
+    # Reconstruction protocol: For zero3 we need to zip the partitions together at boundary of each
+    # param, re-consolidating each param, while dealing with padding if any
+
+    # merge list of dicts, preserving order
+    param_shapes = {k: v for d in param_shapes for k, v in d.items()}
+
+    if debug:
+        for i in range(world_size):
+            print(f"{FP32_FLAT_GROUPS}[{i}].shape={fp32_flat_groups[i].shape}")
+
+        wanted_params = len(param_shapes)
+        wanted_numel = sum(shape.numel() for shape in param_shapes.values())
+        # not asserting if there is a mismatch due to possible padding
+        avail_numel = fp32_flat_groups[0].numel() * world_size
+        print(f"Trainable params: Have {avail_numel} numels to process.")
+        print(f"Trainable params: Need {wanted_numel} numels in {wanted_params} params.")
+
+    # params
+    # XXX: for huge models that can't fit into the host's RAM we will have to recode this to support
+    # out-of-core computing solution
+    offset = 0
+    total_numel = 0
+    total_params = 0
+    flat_groups_offset = [0] + list(np.cumsum([flat_tensor.numel() for flat_tensor in fp32_flat_groups[0]]))
+    for name, shape in tqdm(param_shapes.items(), desc='Gathering sharded weights'):
+        unpartitioned_numel = shape.numel()
+        total_numel += unpartitioned_numel
+        total_params += 1
+        partitioned_numel, partitioned_padding_numel = zero3_partitioned_param_info(unpartitioned_numel, world_size)
+
+        if debug:
+            print(
+                f"Trainable params: {total_params} {name} full shape: {shape} partition0 numel={partitioned_numel} partitioned_padding_numel={partitioned_padding_numel}"
+            )
+
+        # memory efficient tensor
+        tensor = GatheredTensor(fp32_flat_groups, flat_groups_offset, offset, partitioned_numel, shape)
+        state_dict[name] = tensor
+        offset += partitioned_numel
+
+    offset *= world_size
+
+    # Sanity check
+    if offset != avail_numel:
+        raise ValueError(f"consumed {offset} numels out of {avail_numel} - something is wrong")
+
+    print(f"Reconstructed Trainable fp32 state dict with {total_params} params {total_numel} elements")
+
+
+def _get_fp32_state_dict_from_zero3_checkpoint(world_size, fp32_flat_groups, zero_model_states,
+                                               exclude_frozen_parameters):
+    state_dict = OrderedDict()
+
+    # buffers
+    buffers = zero_model_states[0].buffers
+    state_dict.update(buffers)
+    if debug:
+        print(f"added {len(buffers)} buffers")
+
+    if not exclude_frozen_parameters:
+        _zero3_merge_frozen_params(state_dict, world_size, zero_model_states)
+
+    _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero_model_states)
+
+    # recover shared parameters
+    for pair in zero_model_states[0].shared_params:
+        if pair[1] in state_dict:
+            state_dict[pair[0]] = state_dict[pair[1]]
+
+    return state_dict
+
+
+def to_torch_tensor(state_dict, return_empty_tensor=False):
+    """
+    Convert state_dict of GatheredTensor to torch tensor
+    """
+    torch_state_dict = {}
+    converted_tensors = {}
+    for name, tensor in state_dict.items():
+        tensor_id = id(tensor)
+        if tensor_id in converted_tensors:  # shared tensors
+            shared_tensor = torch_state_dict[converted_tensors[tensor_id]]
+            torch_state_dict[name] = shared_tensor
+        else:
+            converted_tensors[tensor_id] = name
+            if return_empty_tensor:
+                torch_state_dict[name] = torch.empty(tensor.shape, dtype=tensor.dtype)
+            else:
+                torch_state_dict[name] = tensor.contiguous()
+    return torch_state_dict
+
+
+def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                             tag=None,
+                                             exclude_frozen_parameters=False,
+                                             lazy_mode=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with
+    ``load_state_dict()`` and used for training without DeepSpeed or shared with others, for example
+    via a model hub.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in 'latest' file. e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+        - ``lazy_mode``: get state_dict in lazy mode. It returns a dict of pesduo tensor instead of torch tensor, which is more memory efficient.
+          Convert the pesduo tensor to torch tensor by ``.contiguous()``
+
+    Returns:
+        - pytorch ``state_dict``
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        # do the training and checkpoint saving
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir) # already on cpu
+        model = model.cpu() # move to cpu
+        model.load_state_dict(state_dict)
+        # submit to model hub or save the model to share with others
+
+    In this example the ``model`` will no longer be usable in the deepspeed context of the same
+    application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    If you want it all done for you, use ``load_state_dict_from_zero_checkpoint`` instead.
+
+    Note: the above usage may not work if your application doesn't have sufficient free CPU memory.
+    You may need to use the offline approach using the ``zero_to_fp32.py`` script that is saved with
+    the checkpoint. Or you can load state_dict in lazy mode ::
+
+        from deepspeed.utils.zero_to_fp32 import get_fp32_state_dict_from_zero_checkpoint
+        state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, lazy_mode=True) # not on cpu
+        for name, lazy_tensor in state_dict.item():
+            tensor = lazy_tensor.contiguous()  # to cpu
+            print(name, tensor)
+            # del tensor to release memory if it no longer in use
+    """
+    if tag is None:
+        latest_path = os.path.join(checkpoint_dir, 'latest')
+        if os.path.isfile(latest_path):
+            with open(latest_path, 'r') as fd:
+                tag = fd.read().strip()
+        else:
+            raise ValueError(f"Unable to find 'latest' file at {latest_path}")
+
+    ds_checkpoint_dir = os.path.join(checkpoint_dir, tag)
+
+    if not os.path.isdir(ds_checkpoint_dir):
+        raise FileNotFoundError(f"Directory '{ds_checkpoint_dir}' doesn't exist")
+
+    state_dict = _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir, exclude_frozen_parameters)
+    if lazy_mode:
+        return state_dict
+    else:
+        return to_torch_tensor(state_dict)
+
+
+def convert_zero_checkpoint_to_fp32_state_dict(checkpoint_dir,
+                                               output_dir,
+                                               max_shard_size="5GB",
+                                               safe_serialization=False,
+                                               tag=None,
+                                               exclude_frozen_parameters=False):
+    """
+    Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict`` file that can be
+    loaded with ``torch.load(file)`` + ``load_state_dict()`` and used for training without DeepSpeed.
+
+    Args:
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``output_dir``: directory to the pytorch fp32 state_dict output files
+        - ``max_shard_size``: the maximum size for a checkpoint before being sharded, default value is 5GB
+        - ``safe_serialization``:  whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+        - ``exclude_frozen_parameters``: exclude frozen parameters
+    """
+
+    # Dependency pre-check
+    if safe_serialization:
+        try:
+            from safetensors.torch import save_file
+        except ImportError:
+            print('If you want to use `safe_serialization`, please `pip install safetensors`')
+            raise
+    if max_shard_size is not None:
+        try:
+            from huggingface_hub import split_torch_state_dict_into_shards
+        except ImportError:
+            print('If you want to use `max_shard_size`, please `pip install huggingface_hub`')
+            raise
+
+    # Convert zero checkpoint to state_dict
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir,
+                                                          tag,
+                                                          exclude_frozen_parameters,
+                                                          lazy_mode=True)
+
+    # Shard the model if it is too big.
+    weights_name = "model.safetensors" if safe_serialization else "pytorch_model.bin"
+    if max_shard_size is not None:
+        filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(".safetensors", "{suffix}.safetensors")
+        # an memory-efficient approach for sharding
+        empty_state_dict = to_torch_tensor(state_dict, return_empty_tensor=True)
+        state_dict_split = split_torch_state_dict_into_shards(empty_state_dict,
+                                                              filename_pattern=filename_pattern,
+                                                              max_shard_size=max_shard_size)
+    else:
+        from collections import namedtuple
+        StateDictSplit = namedtuple("StateDictSplit", ["is_sharded", "filename_to_tensors"])
+        state_dict_split = StateDictSplit(is_sharded=False,
+                                          filename_to_tensors={weights_name: list(state_dict.keys())})
+
+    # Save the model by shard
+    os.makedirs(output_dir, exist_ok=True)
+    filename_to_tensors = state_dict_split.filename_to_tensors.items()
+    for shard_file, tensors in tqdm(filename_to_tensors, desc="Saving checkpoint shards"):
+        shard_state_dict = {tensor_name: state_dict[tensor_name] for tensor_name in tensors}
+        shard_state_dict = to_torch_tensor(shard_state_dict)
+        output_path = os.path.join(output_dir, shard_file)
+        if safe_serialization:
+            save_file(shard_state_dict, output_path, metadata={"format": "pt"})
+        else:
+            torch.save(shard_state_dict, output_path)
+        # release the memory of current shard
+        for tensor_name in list(shard_state_dict.keys()):
+            del state_dict[tensor_name]
+            del shard_state_dict[tensor_name]
+        del shard_state_dict
+        gc.collect()
+
+    # Save index if sharded
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": state_dict_split.metadata,
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        save_index_file = "model.safetensors.index.json" if safe_serialization else "pytorch_model.bin.index.json"
+        save_index_file = os.path.join(output_dir, save_index_file)
+        with open(save_index_file, "w", encoding="utf-8") as f:
+            content = json.dumps(index, indent=2, sort_keys=True) + "\n"
+            f.write(content)
+
+
+def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
+    """
+    1. Put the provided model to cpu
+    2. Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated ``state_dict``
+    3. Load it into the provided model
+
+    Args:
+        - ``model``: the model object to update
+        - ``checkpoint_dir``: path to the desired checkpoint folder. (one that contains the tag-folder, like ``global_step14``)
+        - ``tag``: checkpoint tag used as a unique identifier for checkpoint. If not provided will attempt to load tag in the file named ``latest`` in the checkpoint folder, e.g., ``global_step14``
+
+    Returns:
+        - ``model`: modified model
+
+    Make sure you have plenty of CPU memory available before you call this function. If you don't
+    have enough use the ``zero_to_fp32.py`` utility to do the conversion. You will find it
+    conveniently placed for you in the checkpoint folder.
+
+    A typical usage might be ::
+
+        from deepspeed.utils.zero_to_fp32 import load_state_dict_from_zero_checkpoint
+        model = load_state_dict_from_zero_checkpoint(trainer.model, checkpoint_dir)
+        # submit to model hub or save the model to share with others
+
+    Note, that once this was run, the ``model`` will no longer be usable in the deepspeed context
+    of the same application. i.e. you will need to re-initialize the deepspeed engine, since
+    ``model.load_state_dict(state_dict)`` will remove all the deepspeed magic from it.
+
+    """
+    logger.info(f"Extracting fp32 weights")
+    state_dict = get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag)
+
+    logger.info(f"Overwriting model with fp32 weights")
+    model = model.cpu()
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("checkpoint_dir",
+                        type=str,
+                        help="path to the desired checkpoint folder, e.g., path/checkpoint-12")
+    parser.add_argument("output_dir",
+                        type=str,
+                        help="directory to the pytorch fp32 state_dict output files"
+                        "(e.g. path/checkpoint-12-output/)")
+    parser.add_argument(
+        "--max_shard_size",
+        type=str,
+        default="5GB",
+        help="The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size"
+        "lower than this size. If expressed as a string, needs to be digits followed by a unit (like `5MB`"
+        "We default it to 5GB in order for models to be able to run easily on free-tier google colab instances"
+        "without CPU OOM issues.")
+    parser.add_argument(
+        "--safe_serialization",
+        default=False,
+        action='store_true',
+        help="Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).")
+    parser.add_argument("-t",
+                        "--tag",
+                        type=str,
+                        default=None,
+                        help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1")
+    parser.add_argument("--exclude_frozen_parameters", action='store_true', help="exclude frozen parameters")
+    parser.add_argument("-d", "--debug", action='store_true', help="enable debug")
+    args = parser.parse_args()
+
+    debug = args.debug
+
+    convert_zero_checkpoint_to_fp32_state_dict(args.checkpoint_dir,
+                                               args.output_dir,
+                                               max_shard_size=args.max_shard_size,
+                                               safe_serialization=args.safe_serialization,
+                                               tag=args.tag,
+                                               exclude_frozen_parameters=args.exclude_frozen_parameters)