Hugging Face's logo

caffeic
/
tinystarcoder-reward-tldr

Summarization
Transformers
Safetensors
gpt_bigcode
text-classification
reward-model
reward-trainer
trl
rlhf
preference-learning
text-generation-inference
Model card Files
xet
 Community
caffeic commited on
Commit
65d4134
·
verified ·
1 Parent(s): 02e653f

Upload folder using huggingface_hub

Browse files
Files changed (24) hide show
  1. README.md +58 -0
  2. checkpoint-1450/config.json +48 -0
  3. checkpoint-1450/model.safetensors +3 -0
  4. checkpoint-1450/optimizer.pt +3 -0
  5. checkpoint-1450/rng_state.pth +3 -0
  6. checkpoint-1450/scheduler.pt +3 -0
  7. checkpoint-1450/tokenizer.json +0 -0
  8. checkpoint-1450/tokenizer_config.json +34 -0
  9. checkpoint-1450/trainer_state.json +2325 -0
  10. checkpoint-1450/training_args.bin +3 -0
  11. checkpoint-1500/config.json +48 -0
  12. checkpoint-1500/model.safetensors +3 -0
  13. checkpoint-1500/optimizer.pt +3 -0
  14. checkpoint-1500/rng_state.pth +3 -0
  15. checkpoint-1500/scheduler.pt +3 -0
  16. checkpoint-1500/tokenizer.json +0 -0
  17. checkpoint-1500/tokenizer_config.json +34 -0
  18. checkpoint-1500/trainer_state.json +2404 -0
  19. checkpoint-1500/training_args.bin +3 -0
  20. config.json +48 -0
  21. model.safetensors +3 -0
  22. tokenizer.json +0 -0
  23. tokenizer_config.json +34 -0
  24. training_args.bin +3 -0
README.md ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model: bigcode/tiny_starcoder_py
3
+ library_name: transformers
4
+ model_name: tinystarcoder-rlhf-model
5
+ tags:
6
+ - generated_from_trainer
7
+ - reward-trainer
8
+ - trl
9
+ licence: license
10
+ ---
11
+
12
+ # Model Card for tinystarcoder-rlhf-model
13
+
14
+ This model is a fine-tuned version of [bigcode/tiny_starcoder_py](https://huggingface.co/bigcode/tiny_starcoder_py).
15
+ It has been trained using [TRL](https://github.com/huggingface/trl).
16
+
17
+ ## Quick start
18
+
19
+ ```python
20
+ from transformers import pipeline
21
+
22
+ text = "The capital of France is Paris."
23
+ rewarder = pipeline(model="None", device="cuda")
24
+ output = rewarder(text)[0]
25
+ print(output["score"])
26
+ ```
27
+
28
+ ## Training procedure
29
+
30
+
31
+
32
+
33
+
34
+ This model was trained with Reward.
35
+
36
+ ### Framework versions
37
+
38
+ - TRL: 1.5.0
39
+ - Transformers: 5.0.0
40
+ - Pytorch: 2.11.0+cu128
41
+ - Datasets: 4.8.5
42
+ - Tokenizers: 0.22.2
43
+
44
+ ## Citations
45
+
46
+
47
+
48
+ Cite TRL as:
49
+
50
+ ```bibtex
51
+ @software{vonwerra2020trl,
52
+ title = {{TRL: Transformers Reinforcement Learning}},
53
+ author = {von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallouédec, Quentin},
54
+ license = {Apache-2.0},
55
+ url = {https://github.com/huggingface/trl},
56
+ year = {2020}
57
+ }
58
+ ```
checkpoint-1450/config.json ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "activation_function": "gelu_pytorch_tanh",
3
+ "add_cross_attention": false,
4
+ "architectures": [
5
+ "GPTBigCodeForSequenceClassification"
6
+ ],
7
+ "attention_softmax_in_fp32": true,
8
+ "attn_pdrop": 0.1,
9
+ "bos_token_id": 0,
10
+ "dtype": "float32",
11
+ "embd_pdrop": 0.1,
12
+ "eos_token_id": 0,
13
+ "id2label": {
14
+ "0": "LABEL_0"
15
+ },
16
+ "inference_runner": 0,
17
+ "initializer_range": 0.02,
18
+ "label2id": {
19
+ "LABEL_0": 0
20
+ },
21
+ "layer_norm_epsilon": 1e-05,
22
+ "max_batch_size": null,
23
+ "max_sequence_length": null,
24
+ "model_type": "gpt_bigcode",
25
+ "multi_query": true,
26
+ "n_embd": 768,
27
+ "n_head": 12,
28
+ "n_inner": 3072,
29
+ "n_layer": 20,
30
+ "n_positions": 8192,
31
+ "num_key_value_heads": 1,
32
+ "pad_key_length": true,
33
+ "pad_token_id": 0,
34
+ "pre_allocate_kv_cache": false,
35
+ "resid_pdrop": 0.1,
36
+ "scale_attention_softmax_in_fp32": true,
37
+ "scale_attn_weights": true,
38
+ "summary_activation": null,
39
+ "summary_first_dropout": 0.1,
40
+ "summary_proj_to_labels": true,
41
+ "summary_type": "cls_index",
42
+ "summary_use_proj": true,
43
+ "tie_word_embeddings": true,
44
+ "transformers_version": "5.0.0",
45
+ "use_cache": false,
46
+ "validate_runner_input": true,
47
+ "vocab_size": 49152
48
+ }
checkpoint-1450/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:97445f91302f050ea2f4ec9321ea358c8d81ee7996478662f980f8ea0e35a6cb
3
+ size 656604448
checkpoint-1450/optimizer.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8ccd5c5c93605b1293e1d7012c320378be7af14a7baaa76f9d5844e394ca5128
3
+ size 1313367755
checkpoint-1450/rng_state.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:add4c6b80e12787b33ec64bd8ef442fe26b25a887918cbcec87d07a4918d6f71
3
+ size 14645
checkpoint-1450/scheduler.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:415992a7281ec20a82057ed8b15e1814ea421e307267fa8d620488aa8dcaeaec
3
+ size 1465
checkpoint-1450/tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
checkpoint-1450/tokenizer_config.json ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "backend": "tokenizers",
4
+ "bos_token": "<|endoftext|>",
5
+ "eos_token": "<|endoftext|>",
6
+ "errors": "replace",
7
+ "extra_special_tokens": [
8
+ "<|endoftext|>",
9
+ "<fim_prefix>",
10
+ "<fim_middle>",
11
+ "<fim_suffix>",
12
+ "<fim_pad>",
13
+ "<filename>",
14
+ "<gh_stars>",
15
+ "<issue_start>",
16
+ "<issue_comment>",
17
+ "<issue_closed>",
18
+ "<jupyter_start>",
19
+ "<jupyter_text>",
20
+ "<jupyter_code>",
21
+ "<jupyter_output>",
22
+ "<empty_output>",
23
+ "<commit_before>",
24
+ "<commit_msg>",
25
+ "<commit_after>",
26
+ "<reponame>"
27
+ ],
28
+ "is_local": false,
29
+ "model_max_length": 1000000000000000019884624838656,
30
+ "pad_token": "<|endoftext|>",
31
+ "tokenizer_class": "GPT2Tokenizer",
32
+ "unk_token": "<|endoftext|>",
33
+ "vocab_size": 49152
34
+ }
checkpoint-1450/trainer_state.json ADDED
@@ -0,0 +1,2325 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_global_step": null,
3
+ "best_metric": null,
4
+ "best_model_checkpoint": null,
5
+ "epoch": 1.9333333333333333,
6
+ "eval_steps": 50,
7
+ "global_step": 1450,
8
+ "is_hyper_param_search": false,
9
+ "is_local_process_zero": true,
10
+ "is_world_process_zero": true,
11
+ "log_history": [
12
+ {
13
+ "accuracy": 0.65,
14
+ "epoch": 0.013333333333333334,
15
+ "grad_norm": 26.781282424926758,
16
+ "learning_rate": 4.970000000000001e-06,
17
+ "loss": 0.6315235137939453,
18
+ "margin": 0.18203125,
19
+ "max_reward": -1.005078125,
20
+ "mean_reward": -1.4798828125,
21
+ "min_reward": -1.972265625,
22
+ "num_tokens": 16214.0,
23
+ "step": 10
24
+ },
25
+ {
26
+ "accuracy": 0.35,
27
+ "epoch": 0.02666666666666667,
28
+ "grad_norm": 18.78548812866211,
29
+ "learning_rate": 4.936666666666667e-06,
30
+ "loss": 0.9058952331542969,
31
+ "margin": -0.21234130859375,
32
+ "max_reward": -0.0454345703125,
33
+ "mean_reward": -1.138067626953125,
34
+ "min_reward": -2.29296875,
35
+ "num_tokens": 32750.0,
36
+ "step": 20
37
+ },
38
+ {
39
+ "accuracy": 0.65,
40
+ "epoch": 0.04,
41
+ "grad_norm": 23.18710708618164,
42
+ "learning_rate": 4.9033333333333335e-06,
43
+ "loss": 0.666615343093872,
44
+ "margin": 0.1455078125,
45
+ "max_reward": -0.2720703125,
46
+ "mean_reward": -0.9361328125,
47
+ "min_reward": -1.67734375,
48
+ "num_tokens": 49136.0,
49
+ "step": 30
50
+ },
51
+ {
52
+ "accuracy": 0.55,
53
+ "epoch": 0.05333333333333334,
54
+ "grad_norm": 23.498512268066406,
55
+ "learning_rate": 4.87e-06,
56
+ "loss": 0.775050687789917,
57
+ "margin": -0.038720703125,
58
+ "max_reward": -0.69931640625,
59
+ "mean_reward": -1.3807861328125,
60
+ "min_reward": -1.86953125,
61
+ "num_tokens": 64999.0,
62
+ "step": 40
63
+ },
64
+ {
65
+ "accuracy": 0.7,
66
+ "epoch": 0.06666666666666667,
67
+ "grad_norm": 25.616416931152344,
68
+ "learning_rate": 4.836666666666667e-06,
69
+ "loss": 0.6898736953735352,
70
+ "margin": 0.26455078125,
71
+ "max_reward": -0.5318359375,
72
+ "mean_reward": -1.313720703125,
73
+ "min_reward": -2.13203125,
74
+ "num_tokens": 82563.0,
75
+ "step": 50
76
+ },
77
+ {
78
+ "epoch": 0.06666666666666667,
79
+ "eval_accuracy": 0.532,
80
+ "eval_loss": 0.7488015294075012,
81
+ "eval_margin": 0.0769696044921875,
82
+ "eval_max_reward": -1.1200030517578126,
83
+ "eval_mean_reward": -1.4668482055664063,
84
+ "eval_min_reward": -1.813693359375,
85
+ "eval_num_tokens": 82563.0,
86
+ "eval_runtime": 59.7497,
87
+ "eval_samples_per_second": 8.368,
88
+ "eval_steps_per_second": 8.368,
89
+ "step": 50
90
+ },
91
+ {
92
+ "accuracy": 0.5,
93
+ "epoch": 0.08,
94
+ "grad_norm": 47.123233795166016,
95
+ "learning_rate": 4.803333333333334e-06,
96
+ "loss": 0.7043657302856445,
97
+ "margin": 0.144775390625,
98
+ "max_reward": -0.82724609375,
99
+ "mean_reward": -1.5813232421875,
100
+ "min_reward": -2.2578125,
101
+ "num_tokens": 100158.0,
102
+ "step": 60
103
+ },
104
+ {
105
+ "accuracy": 0.45,
106
+ "epoch": 0.09333333333333334,
107
+ "grad_norm": 43.35136032104492,
108
+ "learning_rate": 4.77e-06,
109
+ "loss": 0.7232241630554199,
110
+ "margin": 0.070654296875,
111
+ "max_reward": -1.10107421875,
112
+ "mean_reward": -1.6929443359375,
113
+ "min_reward": -2.36640625,
114
+ "num_tokens": 116349.0,
115
+ "step": 70
116
+ },
117
+ {
118
+ "accuracy": 0.35,
119
+ "epoch": 0.10666666666666667,
120
+ "grad_norm": 38.70017623901367,
121
+ "learning_rate": 4.736666666666667e-06,
122
+ "loss": 0.9364012718200684,
123
+ "margin": -0.284765625,
124
+ "max_reward": -0.4794921875,
125
+ "mean_reward": -1.26298828125,
126
+ "min_reward": -1.909375,
127
+ "num_tokens": 132123.0,
128
+ "step": 80
129
+ },
130
+ {
131
+ "accuracy": 0.45,
132
+ "epoch": 0.12,
133
+ "grad_norm": 29.024276733398438,
134
+ "learning_rate": 4.703333333333334e-06,
135
+ "loss": 0.8137557983398438,
136
+ "margin": -0.1231597900390625,
137
+ "max_reward": 0.445391845703125,
138
+ "mean_reward": -0.22428436279296876,
139
+ "min_reward": -0.788671875,
140
+ "num_tokens": 147982.0,
141
+ "step": 90
142
+ },
143
+ {
144
+ "accuracy": 0.45,
145
+ "epoch": 0.13333333333333333,
146
+ "grad_norm": 34.12649154663086,
147
+ "learning_rate": 4.670000000000001e-06,
148
+ "loss": 0.7763675689697266,
149
+ "margin": -0.06761474609375,
150
+ "max_reward": 0.74072265625,
151
+ "mean_reward": 0.255206298828125,
152
+ "min_reward": -0.2246826171875,
153
+ "num_tokens": 164058.0,
154
+ "step": 100
155
+ },
156
+ {
157
+ "epoch": 0.13333333333333333,
158
+ "eval_accuracy": 0.51,
159
+ "eval_loss": 0.7375237941741943,
160
+ "eval_margin": 0.008534343719482421,
161
+ "eval_max_reward": 0.5244478607177734,
162
+ "eval_mean_reward": 0.2890134220123291,
163
+ "eval_min_reward": 0.053578983306884764,
164
+ "eval_num_tokens": 164058.0,
165
+ "eval_runtime": 62.906,
166
+ "eval_samples_per_second": 7.948,
167
+ "eval_steps_per_second": 7.948,
168
+ "step": 100
169
+ },
170
+ {
171
+ "accuracy": 0.45,
172
+ "epoch": 0.14666666666666667,
173
+ "grad_norm": 41.91898727416992,
174
+ "learning_rate": 4.6366666666666674e-06,
175
+ "loss": 0.7357148647308349,
176
+ "margin": -0.05003662109375,
177
+ "max_reward": 0.8333984375,
178
+ "mean_reward": 0.398614501953125,
179
+ "min_reward": -0.0110595703125,
180
+ "num_tokens": 180287.0,
181
+ "step": 110
182
+ },
183
+ {
184
+ "accuracy": 0.45,
185
+ "epoch": 0.16,
186
+ "grad_norm": 27.607086181640625,
187
+ "learning_rate": 4.603333333333334e-06,
188
+ "loss": 0.7425424575805664,
189
+ "margin": -0.05296630859375,
190
+ "max_reward": 0.73359375,
191
+ "mean_reward": 0.207928466796875,
192
+ "min_reward": -0.285498046875,
193
+ "num_tokens": 195395.0,
194
+ "step": 120
195
+ },
196
+ {
197
+ "accuracy": 0.55,
198
+ "epoch": 0.17333333333333334,
199
+ "grad_norm": 18.465009689331055,
200
+ "learning_rate": 4.57e-06,
201
+ "loss": 0.7002963066101074,
202
+ "margin": 0.0166015625,
203
+ "max_reward": 0.15341796875,
204
+ "mean_reward": -0.245654296875,
205
+ "min_reward": -0.6259765625,
206
+ "num_tokens": 211423.0,
207
+ "step": 130
208
+ },
209
+ {
210
+ "accuracy": 0.85,
211
+ "epoch": 0.18666666666666668,
212
+ "grad_norm": 24.311744689941406,
213
+ "learning_rate": 4.536666666666667e-06,
214
+ "loss": 0.5668483257293702,
215
+ "margin": 0.39488706588745115,
216
+ "max_reward": 0.4626953125,
217
+ "mean_reward": -0.23407502174377443,
218
+ "min_reward": -0.8857421875,
219
+ "num_tokens": 227155.0,
220
+ "step": 140
221
+ },
222
+ {
223
+ "accuracy": 0.55,
224
+ "epoch": 0.2,
225
+ "grad_norm": 27.90814971923828,
226
+ "learning_rate": 4.503333333333333e-06,
227
+ "loss": 0.9236076354980469,
228
+ "margin": 0.2251708984375,
229
+ "max_reward": 1.816796875,
230
+ "mean_reward": 0.43135986328125,
231
+ "min_reward": -1.107373046875,
232
+ "num_tokens": 243609.0,
233
+ "step": 150
234
+ },
235
+ {
236
+ "epoch": 0.2,
237
+ "eval_accuracy": 0.568,
238
+ "eval_loss": 0.9139922857284546,
239
+ "eval_margin": 0.21751553344726562,
240
+ "eval_max_reward": 1.1517796630859376,
241
+ "eval_mean_reward": 0.4644862823486328,
242
+ "eval_min_reward": -0.22280709838867188,
243
+ "eval_num_tokens": 243609.0,
244
+ "eval_runtime": 62.3021,
245
+ "eval_samples_per_second": 8.025,
246
+ "eval_steps_per_second": 8.025,
247
+ "step": 150
248
+ },
249
+ {
250
+ "accuracy": 0.5,
251
+ "epoch": 0.21333333333333335,
252
+ "grad_norm": 47.18056869506836,
253
+ "learning_rate": 4.47e-06,
254
+ "loss": 1.0227947235107422,
255
+ "margin": -0.06168212890625,
256
+ "max_reward": 1.634716796875,
257
+ "mean_reward": 0.546588134765625,
258
+ "min_reward": -0.413671875,
259
+ "num_tokens": 261657.0,
260
+ "step": 160
261
+ },
262
+ {
263
+ "accuracy": 0.5,
264
+ "epoch": 0.22666666666666666,
265
+ "grad_norm": 19.242305755615234,
266
+ "learning_rate": 4.436666666666667e-06,
267
+ "loss": 0.6836428642272949,
268
+ "margin": 0.1833953857421875,
269
+ "max_reward": 0.95390625,
270
+ "mean_reward": 0.25375518798828123,
271
+ "min_reward": -0.5401123046875,
272
+ "num_tokens": 277075.0,
273
+ "step": 170
274
+ },
275
+ {
276
+ "accuracy": 0.65,
277
+ "epoch": 0.24,
278
+ "grad_norm": 13.613568305969238,
279
+ "learning_rate": 4.403333333333334e-06,
280
+ "loss": 0.5862503528594971,
281
+ "margin": 0.335546875,
282
+ "max_reward": 0.95908203125,
283
+ "mean_reward": 0.436962890625,
284
+ "min_reward": -0.14921875,
285
+ "num_tokens": 291771.0,
286
+ "step": 180
287
+ },
288
+ {
289
+ "accuracy": 0.75,
290
+ "epoch": 0.25333333333333335,
291
+ "grad_norm": 16.98612403869629,
292
+ "learning_rate": 4.3700000000000005e-06,
293
+ "loss": 0.4962740898132324,
294
+ "margin": 0.72353515625,
295
+ "max_reward": 2.7734375,
296
+ "mean_reward": 1.643115234375,
297
+ "min_reward": 0.7041015625,
298
+ "num_tokens": 306869.0,
299
+ "step": 190
300
+ },
301
+ {
302
+ "accuracy": 0.55,
303
+ "epoch": 0.26666666666666666,
304
+ "grad_norm": 5.9575724601745605,
305
+ "learning_rate": 4.336666666666667e-06,
306
+ "loss": 0.8757614135742188,
307
+ "margin": 0.180908203125,
308
+ "max_reward": 3.7546875,
309
+ "mean_reward": 2.4668212890625,
310
+ "min_reward": 1.08994140625,
311
+ "num_tokens": 322209.0,
312
+ "step": 200
313
+ },
314
+ {
315
+ "epoch": 0.26666666666666666,
316
+ "eval_accuracy": 0.558,
317
+ "eval_loss": 0.8031588196754456,
318
+ "eval_margin": 0.211654052734375,
319
+ "eval_max_reward": 3.3186171875,
320
+ "eval_mean_reward": 2.7612215576171875,
321
+ "eval_min_reward": 2.203825927734375,
322
+ "eval_num_tokens": 322209.0,
323
+ "eval_runtime": 62.4301,
324
+ "eval_samples_per_second": 8.009,
325
+ "eval_steps_per_second": 8.009,
326
+ "step": 200
327
+ },
328
+ {
329
+ "accuracy": 0.5,
330
+ "epoch": 0.28,
331
+ "grad_norm": 20.00104331970215,
332
+ "learning_rate": 4.303333333333334e-06,
333
+ "loss": 0.742474889755249,
334
+ "margin": 0.2575439453125,
335
+ "max_reward": 3.06640625,
336
+ "mean_reward": 1.81966552734375,
337
+ "min_reward": 0.197021484375,
338
+ "num_tokens": 338225.0,
339
+ "step": 210
340
+ },
341
+ {
342
+ "accuracy": 0.5,
343
+ "epoch": 0.29333333333333333,
344
+ "grad_norm": 23.415834426879883,
345
+ "learning_rate": 4.270000000000001e-06,
346
+ "loss": 0.8180597305297852,
347
+ "margin": 0.2842529296875,
348
+ "max_reward": 4.1375,
349
+ "mean_reward": 2.89720458984375,
350
+ "min_reward": 1.491162109375,
351
+ "num_tokens": 352992.0,
352
+ "step": 220
353
+ },
354
+ {
355
+ "accuracy": 0.5,
356
+ "epoch": 0.30666666666666664,
357
+ "grad_norm": 28.938772201538086,
358
+ "learning_rate": 4.236666666666667e-06,
359
+ "loss": 0.8096456527709961,
360
+ "margin": -0.066015625,
361
+ "max_reward": 3.71875,
362
+ "mean_reward": 2.8822265625,
363
+ "min_reward": 1.9859375,
364
+ "num_tokens": 368570.0,
365
+ "step": 230
366
+ },
367
+ {
368
+ "accuracy": 0.65,
369
+ "epoch": 0.32,
370
+ "grad_norm": 19.143875122070312,
371
+ "learning_rate": 4.2033333333333335e-06,
372
+ "loss": 0.5784501552581787,
373
+ "margin": 0.43447265625,
374
+ "max_reward": 2.54921875,
375
+ "mean_reward": 1.823486328125,
376
+ "min_reward": 1.0931640625,
377
+ "num_tokens": 384794.0,
378
+ "step": 240
379
+ },
380
+ {
381
+ "accuracy": 0.6,
382
+ "epoch": 0.3333333333333333,
383
+ "grad_norm": 25.47344207763672,
384
+ "learning_rate": 4.17e-06,
385
+ "loss": 0.7684556484222412,
386
+ "margin": 0.04365234375,
387
+ "max_reward": 3.128125,
388
+ "mean_reward": 2.072509765625,
389
+ "min_reward": 1.1900390625,
390
+ "num_tokens": 401148.0,
391
+ "step": 250
392
+ },
393
+ {
394
+ "epoch": 0.3333333333333333,
395
+ "eval_accuracy": 0.572,
396
+ "eval_loss": 0.713194727897644,
397
+ "eval_margin": 0.1503466796875,
398
+ "eval_max_reward": 2.57871484375,
399
+ "eval_mean_reward": 2.22652490234375,
400
+ "eval_min_reward": 1.8743349609375,
401
+ "eval_num_tokens": 401148.0,
402
+ "eval_runtime": 62.3122,
403
+ "eval_samples_per_second": 8.024,
404
+ "eval_steps_per_second": 8.024,
405
+ "step": 250
406
+ },
407
+ {
408
+ "accuracy": 0.3,
409
+ "epoch": 0.3466666666666667,
410
+ "grad_norm": 27.635066986083984,
411
+ "learning_rate": 4.136666666666667e-06,
412
+ "loss": 0.9010297775268554,
413
+ "margin": -0.2201416015625,
414
+ "max_reward": 2.42265625,
415
+ "mean_reward": 1.72039794921875,
416
+ "min_reward": 0.801123046875,
417
+ "num_tokens": 416243.0,
418
+ "step": 260
419
+ },
420
+ {
421
+ "accuracy": 0.65,
422
+ "epoch": 0.36,
423
+ "grad_norm": 22.39845848083496,
424
+ "learning_rate": 4.1033333333333336e-06,
425
+ "loss": 0.7369675159454345,
426
+ "margin": 0.000390625,
427
+ "max_reward": 2.31796875,
428
+ "mean_reward": 1.5876953125,
429
+ "min_reward": 0.834375,
430
+ "num_tokens": 432425.0,
431
+ "step": 270
432
+ },
433
+ {
434
+ "accuracy": 0.6,
435
+ "epoch": 0.37333333333333335,
436
+ "grad_norm": 30.26738166809082,
437
+ "learning_rate": 4.07e-06,
438
+ "loss": 0.7203866004943847,
439
+ "margin": 0.00186767578125,
440
+ "max_reward": 1.509375,
441
+ "mean_reward": 1.016265869140625,
442
+ "min_reward": 0.5093994140625,
443
+ "num_tokens": 446709.0,
444
+ "step": 280
445
+ },
446
+ {
447
+ "accuracy": 0.6,
448
+ "epoch": 0.38666666666666666,
449
+ "grad_norm": 17.360742568969727,
450
+ "learning_rate": 4.036666666666667e-06,
451
+ "loss": 0.6374303817749023,
452
+ "margin": 0.159130859375,
453
+ "max_reward": 1.0638671875,
454
+ "mean_reward": 0.6469970703125,
455
+ "min_reward": 0.1876953125,
456
+ "num_tokens": 462422.0,
457
+ "step": 290
458
+ },
459
+ {
460
+ "accuracy": 0.45,
461
+ "epoch": 0.4,
462
+ "grad_norm": 24.10837173461914,
463
+ "learning_rate": 4.003333333333334e-06,
464
+ "loss": 0.7569143772125244,
465
+ "margin": 0.0091796875,
466
+ "max_reward": 1.9375,
467
+ "mean_reward": 1.299609375,
468
+ "min_reward": 0.747265625,
469
+ "num_tokens": 480306.0,
470
+ "step": 300
471
+ },
472
+ {
473
+ "epoch": 0.4,
474
+ "eval_accuracy": 0.606,
475
+ "eval_loss": 0.6851577162742615,
476
+ "eval_margin": 0.110994384765625,
477
+ "eval_max_reward": 1.68329541015625,
478
+ "eval_mean_reward": 1.4451339111328125,
479
+ "eval_min_reward": 1.206972412109375,
480
+ "eval_num_tokens": 480306.0,
481
+ "eval_runtime": 62.4657,
482
+ "eval_samples_per_second": 8.004,
483
+ "eval_steps_per_second": 8.004,
484
+ "step": 300
485
+ },
486
+ {
487
+ "accuracy": 0.65,
488
+ "epoch": 0.41333333333333333,
489
+ "grad_norm": 13.338096618652344,
490
+ "learning_rate": 3.97e-06,
491
+ "loss": 0.6534909725189209,
492
+ "margin": 0.14378662109375,
493
+ "max_reward": 1.83671875,
494
+ "mean_reward": 1.229180908203125,
495
+ "min_reward": 0.6009033203125,
496
+ "num_tokens": 495683.0,
497
+ "step": 310
498
+ },
499
+ {
500
+ "accuracy": 0.6,
501
+ "epoch": 0.4266666666666667,
502
+ "grad_norm": 14.492751121520996,
503
+ "learning_rate": 3.936666666666667e-06,
504
+ "loss": 0.6619410991668702,
505
+ "margin": 0.144091796875,
506
+ "max_reward": 2.00234375,
507
+ "mean_reward": 1.4490966796875,
508
+ "min_reward": 0.80947265625,
509
+ "num_tokens": 512780.0,
510
+ "step": 320
511
+ },
512
+ {
513
+ "accuracy": 0.6,
514
+ "epoch": 0.44,
515
+ "grad_norm": 17.708974838256836,
516
+ "learning_rate": 3.903333333333334e-06,
517
+ "loss": 0.6470551013946533,
518
+ "margin": 0.224090576171875,
519
+ "max_reward": 2.8828125,
520
+ "mean_reward": 2.1324859619140626,
521
+ "min_reward": 1.29166259765625,
522
+ "num_tokens": 529065.0,
523
+ "step": 330
524
+ },
525
+ {
526
+ "accuracy": 0.5,
527
+ "epoch": 0.4533333333333333,
528
+ "grad_norm": 18.278581619262695,
529
+ "learning_rate": 3.87e-06,
530
+ "loss": 0.6739425659179688,
531
+ "margin": 0.33515625,
532
+ "max_reward": 3.9375,
533
+ "mean_reward": 2.80546875,
534
+ "min_reward": 1.79609375,
535
+ "num_tokens": 546716.0,
536
+ "step": 340
537
+ },
538
+ {
539
+ "accuracy": 0.45,
540
+ "epoch": 0.4666666666666667,
541
+ "grad_norm": 38.923805236816406,
542
+ "learning_rate": 3.836666666666667e-06,
543
+ "loss": 0.7799544334411621,
544
+ "margin": -0.078515625,
545
+ "max_reward": 4.6078125,
546
+ "mean_reward": 3.6021484375,
547
+ "min_reward": 2.72265625,
548
+ "num_tokens": 563762.0,
549
+ "step": 350
550
+ },
551
+ {
552
+ "epoch": 0.4666666666666667,
553
+ "eval_accuracy": 0.572,
554
+ "eval_loss": 0.7251431345939636,
555
+ "eval_margin": 0.1519296875,
556
+ "eval_max_reward": 4.2361640625,
557
+ "eval_mean_reward": 3.866208984375,
558
+ "eval_min_reward": 3.49625390625,
559
+ "eval_num_tokens": 563762.0,
560
+ "eval_runtime": 62.2654,
561
+ "eval_samples_per_second": 8.03,
562
+ "eval_steps_per_second": 8.03,
563
+ "step": 350
564
+ },
565
+ {
566
+ "accuracy": 0.8,
567
+ "epoch": 0.48,
568
+ "grad_norm": 25.30300521850586,
569
+ "learning_rate": 3.803333333333334e-06,
570
+ "loss": 0.5838129043579101,
571
+ "margin": 0.34296875,
572
+ "max_reward": 4.70625,
573
+ "mean_reward": 3.857421875,
574
+ "min_reward": 3.0,
575
+ "num_tokens": 579498.0,
576
+ "step": 360
577
+ },
578
+ {
579
+ "accuracy": 0.5,
580
+ "epoch": 0.49333333333333335,
581
+ "grad_norm": 41.584007263183594,
582
+ "learning_rate": 3.7700000000000003e-06,
583
+ "loss": 0.8219932556152344,
584
+ "margin": 0.094140625,
585
+ "max_reward": 4.8875,
586
+ "mean_reward": 4.0658203125,
587
+ "min_reward": 3.16328125,
588
+ "num_tokens": 597098.0,
589
+ "step": 370
590
+ },
591
+ {
592
+ "accuracy": 0.65,
593
+ "epoch": 0.5066666666666667,
594
+ "grad_norm": 16.238313674926758,
595
+ "learning_rate": 3.7366666666666667e-06,
596
+ "loss": 0.6716785430908203,
597
+ "margin": 0.2427490234375,
598
+ "max_reward": 3.34375,
599
+ "mean_reward": 2.30460205078125,
600
+ "min_reward": 1.168408203125,
601
+ "num_tokens": 613640.0,
602
+ "step": 380
603
+ },
604
+ {
605
+ "accuracy": 0.6,
606
+ "epoch": 0.52,
607
+ "grad_norm": 25.986833572387695,
608
+ "learning_rate": 3.7033333333333336e-06,
609
+ "loss": 0.716916561126709,
610
+ "margin": 0.42451171875,
611
+ "max_reward": 2.96328125,
612
+ "mean_reward": 1.55986328125,
613
+ "min_reward": 0.221484375,
614
+ "num_tokens": 629441.0,
615
+ "step": 390
616
+ },
617
+ {
618
+ "accuracy": 0.75,
619
+ "epoch": 0.5333333333333333,
620
+ "grad_norm": 23.42198371887207,
621
+ "learning_rate": 3.6700000000000004e-06,
622
+ "loss": 0.6442339897155762,
623
+ "margin": 0.3801513671875,
624
+ "max_reward": 2.6359375,
625
+ "mean_reward": 1.87711181640625,
626
+ "min_reward": 1.043212890625,
627
+ "num_tokens": 645440.0,
628
+ "step": 400
629
+ },
630
+ {
631
+ "epoch": 0.5333333333333333,
632
+ "eval_accuracy": 0.618,
633
+ "eval_loss": 0.7101489305496216,
634
+ "eval_margin": 0.2907686767578125,
635
+ "eval_max_reward": 2.299792724609375,
636
+ "eval_mean_reward": 1.8211738891601563,
637
+ "eval_min_reward": 1.3425550537109374,
638
+ "eval_num_tokens": 645440.0,
639
+ "eval_runtime": 62.216,
640
+ "eval_samples_per_second": 8.037,
641
+ "eval_steps_per_second": 8.037,
642
+ "step": 400
643
+ },
644
+ {
645
+ "accuracy": 0.6,
646
+ "epoch": 0.5466666666666666,
647
+ "grad_norm": 20.917293548583984,
648
+ "learning_rate": 3.6366666666666673e-06,
649
+ "loss": 0.8052460670471191,
650
+ "margin": 0.2086181640625,
651
+ "max_reward": 2.9359375,
652
+ "mean_reward": 1.69215087890625,
653
+ "min_reward": 0.44091796875,
654
+ "num_tokens": 663799.0,
655
+ "step": 410
656
+ },
657
+ {
658
+ "accuracy": 0.7,
659
+ "epoch": 0.56,
660
+ "grad_norm": 16.333770751953125,
661
+ "learning_rate": 3.6033333333333337e-06,
662
+ "loss": 0.6448392868041992,
663
+ "margin": 0.210369873046875,
664
+ "max_reward": 0.69697265625,
665
+ "mean_reward": 0.0324127197265625,
666
+ "min_reward": -0.758642578125,
667
+ "num_tokens": 678669.0,
668
+ "step": 420
669
+ },
670
+ {
671
+ "accuracy": 0.55,
672
+ "epoch": 0.5733333333333334,
673
+ "grad_norm": 29.64971923828125,
674
+ "learning_rate": 3.57e-06,
675
+ "loss": 0.6741883277893066,
676
+ "margin": 0.1592041015625,
677
+ "max_reward": 0.2900146484375,
678
+ "mean_reward": -0.4987548828125,
679
+ "min_reward": -1.176171875,
680
+ "num_tokens": 694858.0,
681
+ "step": 430
682
+ },
683
+ {
684
+ "accuracy": 0.45,
685
+ "epoch": 0.5866666666666667,
686
+ "grad_norm": 26.162715911865234,
687
+ "learning_rate": 3.536666666666667e-06,
688
+ "loss": 0.7661936283111572,
689
+ "margin": -0.02214813232421875,
690
+ "max_reward": -0.0858306884765625,
691
+ "mean_reward": -0.7683937072753906,
692
+ "min_reward": -1.260546875,
693
+ "num_tokens": 710995.0,
694
+ "step": 440
695
+ },
696
+ {
697
+ "accuracy": 0.75,
698
+ "epoch": 0.6,
699
+ "grad_norm": 22.932537078857422,
700
+ "learning_rate": 3.5033333333333334e-06,
701
+ "loss": 0.6263980865478516,
702
+ "margin": 0.1980621337890625,
703
+ "max_reward": -0.37247314453125,
704
+ "mean_reward": -0.7766494750976562,
705
+ "min_reward": -1.234375,
706
+ "num_tokens": 726473.0,
707
+ "step": 450
708
+ },
709
+ {
710
+ "epoch": 0.6,
711
+ "eval_accuracy": 0.628,
712
+ "eval_loss": 0.6893447041511536,
713
+ "eval_margin": 0.17741348266601562,
714
+ "eval_max_reward": -0.00065301513671875,
715
+ "eval_mean_reward": -0.32534046936035155,
716
+ "eval_min_reward": -0.6500279235839844,
717
+ "eval_num_tokens": 726473.0,
718
+ "eval_runtime": 62.4229,
719
+ "eval_samples_per_second": 8.01,
720
+ "eval_steps_per_second": 8.01,
721
+ "step": 450
722
+ },
723
+ {
724
+ "accuracy": 0.45,
725
+ "epoch": 0.6133333333333333,
726
+ "grad_norm": 23.200468063354492,
727
+ "learning_rate": 3.4700000000000002e-06,
728
+ "loss": 0.7532081127166748,
729
+ "margin": -0.0011383056640625,
730
+ "max_reward": 0.3897216796875,
731
+ "mean_reward": -0.43083343505859373,
732
+ "min_reward": -1.10458984375,
733
+ "num_tokens": 743418.0,
734
+ "step": 460
735
+ },
736
+ {
737
+ "accuracy": 0.65,
738
+ "epoch": 0.6266666666666667,
739
+ "grad_norm": 23.413482666015625,
740
+ "learning_rate": 3.436666666666667e-06,
741
+ "loss": 0.6554159641265869,
742
+ "margin": 0.20365447998046876,
743
+ "max_reward": -0.0306304931640625,
744
+ "mean_reward": -0.6653846740722656,
745
+ "min_reward": -1.2474609375,
746
+ "num_tokens": 758877.0,
747
+ "step": 470
748
+ },
749
+ {
750
+ "accuracy": 0.45,
751
+ "epoch": 0.64,
752
+ "grad_norm": 23.571287155151367,
753
+ "learning_rate": 3.4033333333333335e-06,
754
+ "loss": 0.7563723564147949,
755
+ "margin": 0.1056365966796875,
756
+ "max_reward": 1.4620849609375,
757
+ "mean_reward": 0.06235504150390625,
758
+ "min_reward": -1.196258544921875,
759
+ "num_tokens": 775285.0,
760
+ "step": 480
761
+ },
762
+ {
763
+ "accuracy": 0.7,
764
+ "epoch": 0.6533333333333333,
765
+ "grad_norm": 12.891551971435547,
766
+ "learning_rate": 3.3700000000000003e-06,
767
+ "loss": 0.5850713729858399,
768
+ "margin": 0.73206787109375,
769
+ "max_reward": 2.0690673828125,
770
+ "mean_reward": 0.511407470703125,
771
+ "min_reward": -0.6796875,
772
+ "num_tokens": 790684.0,
773
+ "step": 490
774
+ },
775
+ {
776
+ "accuracy": 0.8,
777
+ "epoch": 0.6666666666666666,
778
+ "grad_norm": 64.09451293945312,
779
+ "learning_rate": 3.3366666666666668e-06,
780
+ "loss": 0.498201847076416,
781
+ "margin": 0.911553955078125,
782
+ "max_reward": 1.70078125,
783
+ "mean_reward": 0.3194183349609375,
784
+ "min_reward": -1.081884765625,
785
+ "num_tokens": 806487.0,
786
+ "step": 500
787
+ },
788
+ {
789
+ "epoch": 0.6666666666666666,
790
+ "eval_accuracy": 0.622,
791
+ "eval_loss": 0.7471536993980408,
792
+ "eval_margin": 0.40328579330444336,
793
+ "eval_max_reward": 0.7979255027770996,
794
+ "eval_mean_reward": 0.18772327995300292,
795
+ "eval_min_reward": -0.4224789428710938,
796
+ "eval_num_tokens": 806487.0,
797
+ "eval_runtime": 62.2921,
798
+ "eval_samples_per_second": 8.027,
799
+ "eval_steps_per_second": 8.027,
800
+ "step": 500
801
+ },
802
+ {
803
+ "accuracy": 0.65,
804
+ "epoch": 0.68,
805
+ "grad_norm": 122.53633117675781,
806
+ "learning_rate": 3.303333333333333e-06,
807
+ "loss": 0.7432568550109864,
808
+ "margin": 0.4142311096191406,
809
+ "max_reward": 0.76953125,
810
+ "mean_reward": -0.3576038360595703,
811
+ "min_reward": -1.263671875,
812
+ "num_tokens": 823462.0,
813
+ "step": 510
814
+ },
815
+ {
816
+ "accuracy": 0.75,
817
+ "epoch": 0.6933333333333334,
818
+ "grad_norm": 32.80317306518555,
819
+ "learning_rate": 3.2700000000000005e-06,
820
+ "loss": 0.41660318374633787,
821
+ "margin": 1.13740234375,
822
+ "max_reward": 2.3439453125,
823
+ "mean_reward": 0.789794921875,
824
+ "min_reward": -1.080078125,
825
+ "num_tokens": 838460.0,
826
+ "step": 520
827
+ },
828
+ {
829
+ "accuracy": 0.75,
830
+ "epoch": 0.7066666666666667,
831
+ "grad_norm": 17.606361389160156,
832
+ "learning_rate": 3.236666666666667e-06,
833
+ "loss": 0.6934970378875732,
834
+ "margin": 0.973486328125,
835
+ "max_reward": 4.0046875,
836
+ "mean_reward": 2.1105224609375,
837
+ "min_reward": -0.177734375,
838
+ "num_tokens": 854813.0,
839
+ "step": 530
840
+ },
841
+ {
842
+ "accuracy": 0.55,
843
+ "epoch": 0.72,
844
+ "grad_norm": 156.23760986328125,
845
+ "learning_rate": 3.2033333333333337e-06,
846
+ "loss": 1.1027880668640138,
847
+ "margin": 0.20673828125,
848
+ "max_reward": 5.353125,
849
+ "mean_reward": 3.049951171875,
850
+ "min_reward": 0.3279296875,
851
+ "num_tokens": 870996.0,
852
+ "step": 540
853
+ },
854
+ {
855
+ "accuracy": 0.45,
856
+ "epoch": 0.7333333333333333,
857
+ "grad_norm": 34.310646057128906,
858
+ "learning_rate": 3.17e-06,
859
+ "loss": 1.0296151161193847,
860
+ "margin": 0.0447265625,
861
+ "max_reward": 4.716796875,
862
+ "mean_reward": 2.67509765625,
863
+ "min_reward": 0.359375,
864
+ "num_tokens": 886715.0,
865
+ "step": 550
866
+ },
867
+ {
868
+ "epoch": 0.7333333333333333,
869
+ "eval_accuracy": 0.598,
870
+ "eval_loss": 0.8312986493110657,
871
+ "eval_margin": 0.4697259521484375,
872
+ "eval_max_reward": 3.234610900878906,
873
+ "eval_mean_reward": 2.4688244018554686,
874
+ "eval_min_reward": 1.7030379028320313,
875
+ "eval_num_tokens": 886715.0,
876
+ "eval_runtime": 62.3011,
877
+ "eval_samples_per_second": 8.026,
878
+ "eval_steps_per_second": 8.026,
879
+ "step": 550
880
+ },
881
+ {
882
+ "accuracy": 0.65,
883
+ "epoch": 0.7466666666666667,
884
+ "grad_norm": 2.1129918098449707,
885
+ "learning_rate": 3.1366666666666666e-06,
886
+ "loss": 0.6547252178192139,
887
+ "margin": 0.72010498046875,
888
+ "max_reward": 3.6328125,
889
+ "mean_reward": 1.748907470703125,
890
+ "min_reward": -0.1922607421875,
891
+ "num_tokens": 901702.0,
892
+ "step": 560
893
+ },
894
+ {
895
+ "accuracy": 0.55,
896
+ "epoch": 0.76,
897
+ "grad_norm": 19.6908016204834,
898
+ "learning_rate": 3.103333333333334e-06,
899
+ "loss": 0.6725973129272461,
900
+ "margin": 0.536669921875,
901
+ "max_reward": 0.89931640625,
902
+ "mean_reward": -0.4901123046875,
903
+ "min_reward": -1.93125,
904
+ "num_tokens": 919528.0,
905
+ "step": 570
906
+ },
907
+ {
908
+ "accuracy": 0.5,
909
+ "epoch": 0.7733333333333333,
910
+ "grad_norm": 56.94779586791992,
911
+ "learning_rate": 3.0700000000000003e-06,
912
+ "loss": 0.6944450378417969,
913
+ "margin": 0.448974609375,
914
+ "max_reward": 0.27529296875,
915
+ "mean_reward": -1.3425048828125,
916
+ "min_reward": -2.58046875,
917
+ "num_tokens": 936620.0,
918
+ "step": 580
919
+ },
920
+ {
921
+ "accuracy": 0.85,
922
+ "epoch": 0.7866666666666666,
923
+ "grad_norm": 95.96953582763672,
924
+ "learning_rate": 3.0366666666666667e-06,
925
+ "loss": 0.3615936040878296,
926
+ "margin": 1.50234375,
927
+ "max_reward": 1.0953125,
928
+ "mean_reward": -1.33935546875,
929
+ "min_reward": -3.105078125,
930
+ "num_tokens": 953455.0,
931
+ "step": 590
932
+ },
933
+ {
934
+ "accuracy": 0.7,
935
+ "epoch": 0.8,
936
+ "grad_norm": 88.94745635986328,
937
+ "learning_rate": 3.0033333333333335e-06,
938
+ "loss": 0.9391720771789551,
939
+ "margin": 1.326025390625,
940
+ "max_reward": 5.046875,
941
+ "mean_reward": 2.1633544921875,
942
+ "min_reward": -1.42548828125,
943
+ "num_tokens": 968502.0,
944
+ "step": 600
945
+ },
946
+ {
947
+ "epoch": 0.8,
948
+ "eval_accuracy": 0.598,
949
+ "eval_loss": 1.15407395362854,
950
+ "eval_margin": 0.8387359619140625,
951
+ "eval_max_reward": 3.26913037109375,
952
+ "eval_mean_reward": 1.9263014526367188,
953
+ "eval_min_reward": 0.5834725341796875,
954
+ "eval_num_tokens": 968502.0,
955
+ "eval_runtime": 62.2605,
956
+ "eval_samples_per_second": 8.031,
957
+ "eval_steps_per_second": 8.031,
958
+ "step": 600
959
+ },
960
+ {
961
+ "accuracy": 0.5,
962
+ "epoch": 0.8133333333333334,
963
+ "grad_norm": 32.30426025390625,
964
+ "learning_rate": 2.97e-06,
965
+ "loss": 1.0049471855163574,
966
+ "margin": 0.078076171875,
967
+ "max_reward": 4.67255859375,
968
+ "mean_reward": 1.0933349609375,
969
+ "min_reward": -2.70625,
970
+ "num_tokens": 985736.0,
971
+ "step": 610
972
+ },
973
+ {
974
+ "accuracy": 0.7,
975
+ "epoch": 0.8266666666666667,
976
+ "grad_norm": 51.47792434692383,
977
+ "learning_rate": 2.936666666666667e-06,
978
+ "loss": 0.7427908420562744,
979
+ "margin": 0.39453125,
980
+ "max_reward": 5.634375,
981
+ "mean_reward": 4.305078125,
982
+ "min_reward": 2.32734375,
983
+ "num_tokens": 1002008.0,
984
+ "step": 620
985
+ },
986
+ {
987
+ "accuracy": 0.5,
988
+ "epoch": 0.84,
989
+ "grad_norm": 23.034889221191406,
990
+ "learning_rate": 2.9033333333333336e-06,
991
+ "loss": 0.8153639793395996,
992
+ "margin": -0.12333984375,
993
+ "max_reward": 5.821875,
994
+ "mean_reward": 4.691064453125,
995
+ "min_reward": 3.4435546875,
996
+ "num_tokens": 1019028.0,
997
+ "step": 630
998
+ },
999
+ {
1000
+ "accuracy": 0.35,
1001
+ "epoch": 0.8533333333333334,
1002
+ "grad_norm": 161.87069702148438,
1003
+ "learning_rate": 2.87e-06,
1004
+ "loss": 0.8048789978027344,
1005
+ "margin": -0.009716796875,
1006
+ "max_reward": 5.628125,
1007
+ "mean_reward": 4.9048583984375,
1008
+ "min_reward": 3.97490234375,
1009
+ "num_tokens": 1033366.0,
1010
+ "step": 640
1011
+ },
1012
+ {
1013
+ "accuracy": 0.55,
1014
+ "epoch": 0.8666666666666667,
1015
+ "grad_norm": 15.923206329345703,
1016
+ "learning_rate": 2.836666666666667e-06,
1017
+ "loss": 0.7137910842895507,
1018
+ "margin": 0.00390625,
1019
+ "max_reward": 5.36875,
1020
+ "mean_reward": 4.934765625,
1021
+ "min_reward": 4.5046875,
1022
+ "num_tokens": 1050252.0,
1023
+ "step": 650
1024
+ },
1025
+ {
1026
+ "epoch": 0.8666666666666667,
1027
+ "eval_accuracy": 0.526,
1028
+ "eval_loss": 0.7010688185691833,
1029
+ "eval_margin": 0.0384140625,
1030
+ "eval_max_reward": 5.014875,
1031
+ "eval_mean_reward": 4.85937109375,
1032
+ "eval_min_reward": 4.7038671875,
1033
+ "eval_num_tokens": 1050252.0,
1034
+ "eval_runtime": 62.339,
1035
+ "eval_samples_per_second": 8.021,
1036
+ "eval_steps_per_second": 8.021,
1037
+ "step": 650
1038
+ },
1039
+ {
1040
+ "accuracy": 0.6,
1041
+ "epoch": 0.88,
1042
+ "grad_norm": 15.554156303405762,
1043
+ "learning_rate": 2.8033333333333333e-06,
1044
+ "loss": 0.65819411277771,
1045
+ "margin": 0.16484375,
1046
+ "max_reward": 5.25,
1047
+ "mean_reward": 4.766015625,
1048
+ "min_reward": 4.1578125,
1049
+ "num_tokens": 1068961.0,
1050
+ "step": 660
1051
+ },
1052
+ {
1053
+ "accuracy": 0.55,
1054
+ "epoch": 0.8933333333333333,
1055
+ "grad_norm": 14.689868927001953,
1056
+ "learning_rate": 2.7700000000000006e-06,
1057
+ "loss": 0.68900465965271,
1058
+ "margin": 0.12265625,
1059
+ "max_reward": 4.746875,
1060
+ "mean_reward": 4.285546875,
1061
+ "min_reward": 3.6515625,
1062
+ "num_tokens": 1084178.0,
1063
+ "step": 670
1064
+ },
1065
+ {
1066
+ "accuracy": 0.65,
1067
+ "epoch": 0.9066666666666666,
1068
+ "grad_norm": 30.876516342163086,
1069
+ "learning_rate": 2.736666666666667e-06,
1070
+ "loss": 0.6992377281188965,
1071
+ "margin": 0.0375,
1072
+ "max_reward": 4.3859375,
1073
+ "mean_reward": 4.02265625,
1074
+ "min_reward": 3.6703125,
1075
+ "num_tokens": 1100910.0,
1076
+ "step": 680
1077
+ },
1078
+ {
1079
+ "accuracy": 0.65,
1080
+ "epoch": 0.92,
1081
+ "grad_norm": 24.66133689880371,
1082
+ "learning_rate": 2.7033333333333334e-06,
1083
+ "loss": 0.6626120567321777,
1084
+ "margin": 0.20625,
1085
+ "max_reward": 4.084375,
1086
+ "mean_reward": 3.6171875,
1087
+ "min_reward": 2.8875,
1088
+ "num_tokens": 1116440.0,
1089
+ "step": 690
1090
+ },
1091
+ {
1092
+ "accuracy": 0.45,
1093
+ "epoch": 0.9333333333333333,
1094
+ "grad_norm": 15.120383262634277,
1095
+ "learning_rate": 2.6700000000000003e-06,
1096
+ "loss": 0.696980619430542,
1097
+ "margin": 0.13203125,
1098
+ "max_reward": 4.3109375,
1099
+ "mean_reward": 3.649609375,
1100
+ "min_reward": 3.0734375,
1101
+ "num_tokens": 1132429.0,
1102
+ "step": 700
1103
+ },
1104
+ {
1105
+ "epoch": 0.9333333333333333,
1106
+ "eval_accuracy": 0.582,
1107
+ "eval_loss": 0.7070507407188416,
1108
+ "eval_margin": 0.0835546875,
1109
+ "eval_max_reward": 3.954890625,
1110
+ "eval_mean_reward": 3.70844921875,
1111
+ "eval_min_reward": 3.4620078125,
1112
+ "eval_num_tokens": 1132429.0,
1113
+ "eval_runtime": 62.4281,
1114
+ "eval_samples_per_second": 8.009,
1115
+ "eval_steps_per_second": 8.009,
1116
+ "step": 700
1117
+ },
1118
+ {
1119
+ "accuracy": 0.55,
1120
+ "epoch": 0.9466666666666667,
1121
+ "grad_norm": 14.82724380493164,
1122
+ "learning_rate": 2.6366666666666667e-06,
1123
+ "loss": 0.6956966400146485,
1124
+ "margin": 0.065625,
1125
+ "max_reward": 4.340625,
1126
+ "mean_reward": 3.9,
1127
+ "min_reward": 3.3796875,
1128
+ "num_tokens": 1148620.0,
1129
+ "step": 710
1130
+ },
1131
+ {
1132
+ "accuracy": 0.35,
1133
+ "epoch": 0.96,
1134
+ "grad_norm": 28.879314422607422,
1135
+ "learning_rate": 2.603333333333334e-06,
1136
+ "loss": 0.7410328388214111,
1137
+ "margin": -0.046875,
1138
+ "max_reward": 4.2875,
1139
+ "mean_reward": 3.8375,
1140
+ "min_reward": 3.3515625,
1141
+ "num_tokens": 1164601.0,
1142
+ "step": 720
1143
+ },
1144
+ {
1145
+ "accuracy": 0.65,
1146
+ "epoch": 0.9733333333333334,
1147
+ "grad_norm": 17.8446044921875,
1148
+ "learning_rate": 2.5700000000000004e-06,
1149
+ "loss": 0.6587936878204346,
1150
+ "margin": 0.2615234375,
1151
+ "max_reward": 4.5078125,
1152
+ "mean_reward": 3.79892578125,
1153
+ "min_reward": 3.089453125,
1154
+ "num_tokens": 1179554.0,
1155
+ "step": 730
1156
+ },
1157
+ {
1158
+ "accuracy": 0.55,
1159
+ "epoch": 0.9866666666666667,
1160
+ "grad_norm": 16.74919891357422,
1161
+ "learning_rate": 2.536666666666667e-06,
1162
+ "loss": 0.9084688186645508,
1163
+ "margin": -0.1103515625,
1164
+ "max_reward": 3.65,
1165
+ "mean_reward": 2.99423828125,
1166
+ "min_reward": 1.926953125,
1167
+ "num_tokens": 1195986.0,
1168
+ "step": 740
1169
+ },
1170
+ {
1171
+ "accuracy": 0.6,
1172
+ "epoch": 1.0,
1173
+ "grad_norm": 20.36426544189453,
1174
+ "learning_rate": 2.5033333333333336e-06,
1175
+ "loss": 0.6503221035003662,
1176
+ "margin": 0.196875,
1177
+ "max_reward": 3.8296875,
1178
+ "mean_reward": 3.22890625,
1179
+ "min_reward": 2.6796875,
1180
+ "num_tokens": 1212807.0,
1181
+ "step": 750
1182
+ },
1183
+ {
1184
+ "epoch": 1.0,
1185
+ "eval_accuracy": 0.574,
1186
+ "eval_loss": 0.724712073802948,
1187
+ "eval_margin": 0.1419578857421875,
1188
+ "eval_max_reward": 3.170109375,
1189
+ "eval_mean_reward": 2.827115661621094,
1190
+ "eval_min_reward": 2.4841219482421875,
1191
+ "eval_num_tokens": 1212807.0,
1192
+ "eval_runtime": 62.4609,
1193
+ "eval_samples_per_second": 8.005,
1194
+ "eval_steps_per_second": 8.005,
1195
+ "step": 750
1196
+ },
1197
+ {
1198
+ "accuracy": 0.8,
1199
+ "epoch": 1.0133333333333334,
1200
+ "grad_norm": 25.44012451171875,
1201
+ "learning_rate": 2.47e-06,
1202
+ "loss": 0.5082518577575683,
1203
+ "margin": 0.634228515625,
1204
+ "max_reward": 3.65,
1205
+ "mean_reward": 2.5745849609375,
1206
+ "min_reward": 1.33837890625,
1207
+ "num_tokens": 1230748.0,
1208
+ "step": 760
1209
+ },
1210
+ {
1211
+ "accuracy": 0.6,
1212
+ "epoch": 1.0266666666666666,
1213
+ "grad_norm": 18.013717651367188,
1214
+ "learning_rate": 2.436666666666667e-06,
1215
+ "loss": 0.5603197574615478,
1216
+ "margin": 0.5654541015625,
1217
+ "max_reward": 3.029833984375,
1218
+ "mean_reward": 1.75579833984375,
1219
+ "min_reward": 0.37646484375,
1220
+ "num_tokens": 1245628.0,
1221
+ "step": 770
1222
+ },
1223
+ {
1224
+ "accuracy": 0.55,
1225
+ "epoch": 1.04,
1226
+ "grad_norm": 99.6661148071289,
1227
+ "learning_rate": 2.4033333333333338e-06,
1228
+ "loss": 0.8869245529174805,
1229
+ "margin": 0.205218505859375,
1230
+ "max_reward": 2.080810546875,
1231
+ "mean_reward": -0.0414886474609375,
1232
+ "min_reward": -2.2107421875,
1233
+ "num_tokens": 1262533.0,
1234
+ "step": 780
1235
+ },
1236
+ {
1237
+ "accuracy": 0.7,
1238
+ "epoch": 1.0533333333333332,
1239
+ "grad_norm": 53.62653732299805,
1240
+ "learning_rate": 2.37e-06,
1241
+ "loss": 0.5512317657470703,
1242
+ "margin": 1.0128662109375,
1243
+ "max_reward": 1.4505859375,
1244
+ "mean_reward": -0.33284912109375,
1245
+ "min_reward": -2.0107421875,
1246
+ "num_tokens": 1277096.0,
1247
+ "step": 790
1248
+ },
1249
+ {
1250
+ "accuracy": 0.7,
1251
+ "epoch": 1.0666666666666667,
1252
+ "grad_norm": 95.58021545410156,
1253
+ "learning_rate": 2.3366666666666666e-06,
1254
+ "loss": 0.6172554016113281,
1255
+ "margin": 1.2177459716796875,
1256
+ "max_reward": 1.304345703125,
1257
+ "mean_reward": -1.1034286499023438,
1258
+ "min_reward": -3.26953125,
1259
+ "num_tokens": 1293073.0,
1260
+ "step": 800
1261
+ },
1262
+ {
1263
+ "epoch": 1.0666666666666667,
1264
+ "eval_accuracy": 0.596,
1265
+ "eval_loss": 0.8903509378433228,
1266
+ "eval_margin": 0.564534423828125,
1267
+ "eval_max_reward": -0.50098583984375,
1268
+ "eval_mean_reward": -1.4187645263671875,
1269
+ "eval_min_reward": -2.336543212890625,
1270
+ "eval_num_tokens": 1293073.0,
1271
+ "eval_runtime": 62.3781,
1272
+ "eval_samples_per_second": 8.016,
1273
+ "eval_steps_per_second": 8.016,
1274
+ "step": 800
1275
+ },
1276
+ {
1277
+ "accuracy": 0.65,
1278
+ "epoch": 1.08,
1279
+ "grad_norm": 34.59308624267578,
1280
+ "learning_rate": 2.3033333333333334e-06,
1281
+ "loss": 0.7249170303344726,
1282
+ "margin": 0.968017578125,
1283
+ "max_reward": 0.56064453125,
1284
+ "mean_reward": -1.4824951171875,
1285
+ "min_reward": -3.8390625,
1286
+ "num_tokens": 1308687.0,
1287
+ "step": 810
1288
+ },
1289
+ {
1290
+ "accuracy": 0.8,
1291
+ "epoch": 1.0933333333333333,
1292
+ "grad_norm": 30.92959976196289,
1293
+ "learning_rate": 2.2700000000000003e-06,
1294
+ "loss": 0.39841785430908205,
1295
+ "margin": 1.375030517578125,
1296
+ "max_reward": 1.658984375,
1297
+ "mean_reward": -0.6310943603515625,
1298
+ "min_reward": -3.046484375,
1299
+ "num_tokens": 1324618.0,
1300
+ "step": 820
1301
+ },
1302
+ {
1303
+ "accuracy": 0.7,
1304
+ "epoch": 1.1066666666666667,
1305
+ "grad_norm": 208.02574157714844,
1306
+ "learning_rate": 2.236666666666667e-06,
1307
+ "loss": 0.8557153701782226,
1308
+ "margin": 1.5063720703125,
1309
+ "max_reward": 1.581640625,
1310
+ "mean_reward": -0.71844482421875,
1311
+ "min_reward": -3.421484375,
1312
+ "num_tokens": 1339988.0,
1313
+ "step": 830
1314
+ },
1315
+ {
1316
+ "accuracy": 0.75,
1317
+ "epoch": 1.12,
1318
+ "grad_norm": 279.54534912109375,
1319
+ "learning_rate": 2.2033333333333336e-06,
1320
+ "loss": 0.7527968406677246,
1321
+ "margin": 2.487060546875,
1322
+ "max_reward": 1.5951171875,
1323
+ "mean_reward": -2.3231201171875,
1324
+ "min_reward": -5.828125,
1325
+ "num_tokens": 1356112.0,
1326
+ "step": 840
1327
+ },
1328
+ {
1329
+ "accuracy": 0.7,
1330
+ "epoch": 1.1333333333333333,
1331
+ "grad_norm": 35.76692581176758,
1332
+ "learning_rate": 2.17e-06,
1333
+ "loss": 0.8942340850830078,
1334
+ "margin": 1.6185546875,
1335
+ "max_reward": 0.787890625,
1336
+ "mean_reward": -1.966064453125,
1337
+ "min_reward": -5.40625,
1338
+ "num_tokens": 1372383.0,
1339
+ "step": 850
1340
+ },
1341
+ {
1342
+ "epoch": 1.1333333333333333,
1343
+ "eval_accuracy": 0.596,
1344
+ "eval_loss": 1.0435729026794434,
1345
+ "eval_margin": 0.8098057250976562,
1346
+ "eval_max_reward": 0.7392849731445312,
1347
+ "eval_mean_reward": -0.5019579772949219,
1348
+ "eval_min_reward": -1.743200927734375,
1349
+ "eval_num_tokens": 1372383.0,
1350
+ "eval_runtime": 62.4519,
1351
+ "eval_samples_per_second": 8.006,
1352
+ "eval_steps_per_second": 8.006,
1353
+ "step": 850
1354
+ },
1355
+ {
1356
+ "accuracy": 0.8,
1357
+ "epoch": 1.1466666666666667,
1358
+ "grad_norm": 18.71978759765625,
1359
+ "learning_rate": 2.136666666666667e-06,
1360
+ "loss": 0.6701197147369384,
1361
+ "margin": 1.843017578125,
1362
+ "max_reward": 3.765625,
1363
+ "mean_reward": 0.3057373046875,
1364
+ "min_reward": -3.39443359375,
1365
+ "num_tokens": 1388575.0,
1366
+ "step": 860
1367
+ },
1368
+ {
1369
+ "accuracy": 0.7,
1370
+ "epoch": 1.16,
1371
+ "grad_norm": 150.29055786132812,
1372
+ "learning_rate": 2.1033333333333337e-06,
1373
+ "loss": 0.6811777114868164,
1374
+ "margin": 1.3685546875,
1375
+ "max_reward": 2.4873046875,
1376
+ "mean_reward": 0.19677734375,
1377
+ "min_reward": -2.523046875,
1378
+ "num_tokens": 1403606.0,
1379
+ "step": 870
1380
+ },
1381
+ {
1382
+ "accuracy": 0.75,
1383
+ "epoch": 1.1733333333333333,
1384
+ "grad_norm": 187.19081115722656,
1385
+ "learning_rate": 2.07e-06,
1386
+ "loss": 0.8808813095092773,
1387
+ "margin": 1.917291259765625,
1388
+ "max_reward": 4.8203125,
1389
+ "mean_reward": 2.1001434326171875,
1390
+ "min_reward": -1.4076171875,
1391
+ "num_tokens": 1418488.0,
1392
+ "step": 880
1393
+ },
1394
+ {
1395
+ "accuracy": 0.8,
1396
+ "epoch": 1.1866666666666668,
1397
+ "grad_norm": 44.864192962646484,
1398
+ "learning_rate": 2.036666666666667e-06,
1399
+ "loss": 0.558538818359375,
1400
+ "margin": 2.50665283203125,
1401
+ "max_reward": 4.6734375,
1402
+ "mean_reward": 1.906243896484375,
1403
+ "min_reward": -2.1400634765625,
1404
+ "num_tokens": 1433529.0,
1405
+ "step": 890
1406
+ },
1407
+ {
1408
+ "accuracy": 0.8,
1409
+ "epoch": 1.2,
1410
+ "grad_norm": 46.679744720458984,
1411
+ "learning_rate": 2.0033333333333334e-06,
1412
+ "loss": 0.7218273639678955,
1413
+ "margin": 2.88583984375,
1414
+ "max_reward": 3.57421875,
1415
+ "mean_reward": -0.101611328125,
1416
+ "min_reward": -4.85234375,
1417
+ "num_tokens": 1450110.0,
1418
+ "step": 900
1419
+ },
1420
+ {
1421
+ "epoch": 1.2,
1422
+ "eval_accuracy": 0.608,
1423
+ "eval_loss": 1.363513708114624,
1424
+ "eval_margin": 1.084634536743164,
1425
+ "eval_max_reward": 3.07980419921875,
1426
+ "eval_mean_reward": 1.359497184753418,
1427
+ "eval_min_reward": -0.36080982971191405,
1428
+ "eval_num_tokens": 1450110.0,
1429
+ "eval_runtime": 62.4495,
1430
+ "eval_samples_per_second": 8.006,
1431
+ "eval_steps_per_second": 8.006,
1432
+ "step": 900
1433
+ },
1434
+ {
1435
+ "accuracy": 0.6,
1436
+ "epoch": 1.2133333333333334,
1437
+ "grad_norm": 190.0553741455078,
1438
+ "learning_rate": 1.97e-06,
1439
+ "loss": 0.7698217391967773,
1440
+ "margin": 1.106640625,
1441
+ "max_reward": 5.125,
1442
+ "mean_reward": 1.506640625,
1443
+ "min_reward": -2.765625,
1444
+ "num_tokens": 1466027.0,
1445
+ "step": 910
1446
+ },
1447
+ {
1448
+ "accuracy": 0.6,
1449
+ "epoch": 1.2266666666666666,
1450
+ "grad_norm": 234.94076538085938,
1451
+ "learning_rate": 1.9366666666666666e-06,
1452
+ "loss": 1.6000349044799804,
1453
+ "margin": 1.884661865234375,
1454
+ "max_reward": 1.996875,
1455
+ "mean_reward": -2.3544403076171876,
1456
+ "min_reward": -8.725,
1457
+ "num_tokens": 1481748.0,
1458
+ "step": 920
1459
+ },
1460
+ {
1461
+ "accuracy": 0.55,
1462
+ "epoch": 1.24,
1463
+ "grad_norm": 65.35015869140625,
1464
+ "learning_rate": 1.9033333333333335e-06,
1465
+ "loss": 1.3991120338439942,
1466
+ "margin": 1.012890625,
1467
+ "max_reward": -1.74296875,
1468
+ "mean_reward": -5.5919921875,
1469
+ "min_reward": -9.453125,
1470
+ "num_tokens": 1498020.0,
1471
+ "step": 930
1472
+ },
1473
+ {
1474
+ "accuracy": 0.75,
1475
+ "epoch": 1.2533333333333334,
1476
+ "grad_norm": 57.545166015625,
1477
+ "learning_rate": 1.87e-06,
1478
+ "loss": 0.49834294319152833,
1479
+ "margin": 2.0171875,
1480
+ "max_reward": -3.440625,
1481
+ "mean_reward": -6.62265625,
1482
+ "min_reward": -9.015625,
1483
+ "num_tokens": 1514583.0,
1484
+ "step": 940
1485
+ },
1486
+ {
1487
+ "accuracy": 0.85,
1488
+ "epoch": 1.2666666666666666,
1489
+ "grad_norm": 144.95867919921875,
1490
+ "learning_rate": 1.836666666666667e-06,
1491
+ "loss": 0.5361029624938964,
1492
+ "margin": 3.177197265625,
1493
+ "max_reward": 0.4771484375,
1494
+ "mean_reward": -3.9238525390625,
1495
+ "min_reward": -8.1375,
1496
+ "num_tokens": 1530955.0,
1497
+ "step": 950
1498
+ },
1499
+ {
1500
+ "epoch": 1.2666666666666666,
1501
+ "eval_accuracy": 0.596,
1502
+ "eval_loss": 1.3600239753723145,
1503
+ "eval_margin": 1.1310927734375,
1504
+ "eval_max_reward": -1.742294677734375,
1505
+ "eval_mean_reward": -3.521328125,
1506
+ "eval_min_reward": -5.300361572265625,
1507
+ "eval_num_tokens": 1530955.0,
1508
+ "eval_runtime": 62.6012,
1509
+ "eval_samples_per_second": 7.987,
1510
+ "eval_steps_per_second": 7.987,
1511
+ "step": 950
1512
+ },
1513
+ {
1514
+ "accuracy": 0.5,
1515
+ "epoch": 1.28,
1516
+ "grad_norm": 319.9293518066406,
1517
+ "learning_rate": 1.8033333333333336e-06,
1518
+ "loss": 1.7453046798706056,
1519
+ "margin": 0.3447265625,
1520
+ "max_reward": 1.01591796875,
1521
+ "mean_reward": -3.259765625,
1522
+ "min_reward": -7.2578125,
1523
+ "num_tokens": 1548332.0,
1524
+ "step": 960
1525
+ },
1526
+ {
1527
+ "accuracy": 0.65,
1528
+ "epoch": 1.2933333333333334,
1529
+ "grad_norm": 33.143341064453125,
1530
+ "learning_rate": 1.77e-06,
1531
+ "loss": 0.7040543556213379,
1532
+ "margin": 2.181396484375,
1533
+ "max_reward": -0.30205078125,
1534
+ "mean_reward": -3.7452392578125,
1535
+ "min_reward": -6.275,
1536
+ "num_tokens": 1564625.0,
1537
+ "step": 970
1538
+ },
1539
+ {
1540
+ "accuracy": 0.75,
1541
+ "epoch": 1.3066666666666666,
1542
+ "grad_norm": 31.115459442138672,
1543
+ "learning_rate": 1.7366666666666668e-06,
1544
+ "loss": 0.4173430919647217,
1545
+ "margin": 2.164306640625,
1546
+ "max_reward": 2.28095703125,
1547
+ "mean_reward": -0.9407470703125,
1548
+ "min_reward": -4.04375,
1549
+ "num_tokens": 1580000.0,
1550
+ "step": 980
1551
+ },
1552
+ {
1553
+ "accuracy": 0.8,
1554
+ "epoch": 1.32,
1555
+ "grad_norm": 168.20628356933594,
1556
+ "learning_rate": 1.7033333333333335e-06,
1557
+ "loss": 0.5366679191589355,
1558
+ "margin": 3.0508544921875,
1559
+ "max_reward": 2.8671875,
1560
+ "mean_reward": -1.78248291015625,
1561
+ "min_reward": -5.2830078125,
1562
+ "num_tokens": 1597302.0,
1563
+ "step": 990
1564
+ },
1565
+ {
1566
+ "accuracy": 0.45,
1567
+ "epoch": 1.3333333333333333,
1568
+ "grad_norm": 18.123050689697266,
1569
+ "learning_rate": 1.6700000000000003e-06,
1570
+ "loss": 1.1453736305236817,
1571
+ "margin": 0.8822265625,
1572
+ "max_reward": 4.8984375,
1573
+ "mean_reward": 2.26845703125,
1574
+ "min_reward": -0.678515625,
1575
+ "num_tokens": 1613437.0,
1576
+ "step": 1000
1577
+ },
1578
+ {
1579
+ "epoch": 1.3333333333333333,
1580
+ "eval_accuracy": 0.586,
1581
+ "eval_loss": 1.1093744039535522,
1582
+ "eval_margin": 0.682355224609375,
1583
+ "eval_max_reward": 3.630033935546875,
1584
+ "eval_mean_reward": 2.4088892822265624,
1585
+ "eval_min_reward": 1.18774462890625,
1586
+ "eval_num_tokens": 1613437.0,
1587
+ "eval_runtime": 62.4977,
1588
+ "eval_samples_per_second": 8.0,
1589
+ "eval_steps_per_second": 8.0,
1590
+ "step": 1000
1591
+ },
1592
+ {
1593
+ "accuracy": 0.75,
1594
+ "epoch": 1.3466666666666667,
1595
+ "grad_norm": 35.67937088012695,
1596
+ "learning_rate": 1.6366666666666667e-06,
1597
+ "loss": 0.675491189956665,
1598
+ "margin": 0.5638671875,
1599
+ "max_reward": 4.290625,
1600
+ "mean_reward": 2.44892578125,
1601
+ "min_reward": 0.67578125,
1602
+ "num_tokens": 1629455.0,
1603
+ "step": 1010
1604
+ },
1605
+ {
1606
+ "accuracy": 0.7,
1607
+ "epoch": 1.3599999999999999,
1608
+ "grad_norm": 40.7118034362793,
1609
+ "learning_rate": 1.6033333333333334e-06,
1610
+ "loss": 0.42725315093994143,
1611
+ "margin": 1.8513671875,
1612
+ "max_reward": 4.996875,
1613
+ "mean_reward": 3.25732421875,
1614
+ "min_reward": 0.50625,
1615
+ "num_tokens": 1644785.0,
1616
+ "step": 1020
1617
+ },
1618
+ {
1619
+ "accuracy": 0.8,
1620
+ "epoch": 1.3733333333333333,
1621
+ "grad_norm": 136.8484344482422,
1622
+ "learning_rate": 1.5700000000000002e-06,
1623
+ "loss": 0.6001864433288574,
1624
+ "margin": 3.10028076171875,
1625
+ "max_reward": 4.1550537109375,
1626
+ "mean_reward": 0.757366943359375,
1627
+ "min_reward": -3.25009765625,
1628
+ "num_tokens": 1661544.0,
1629
+ "step": 1030
1630
+ },
1631
+ {
1632
+ "accuracy": 0.75,
1633
+ "epoch": 1.3866666666666667,
1634
+ "grad_norm": 0.3609694242477417,
1635
+ "learning_rate": 1.5366666666666668e-06,
1636
+ "loss": 0.9377962112426758,
1637
+ "margin": 1.72666015625,
1638
+ "max_reward": 3.7875,
1639
+ "mean_reward": 1.237353515625,
1640
+ "min_reward": -2.774609375,
1641
+ "num_tokens": 1677000.0,
1642
+ "step": 1040
1643
+ },
1644
+ {
1645
+ "accuracy": 0.7,
1646
+ "epoch": 1.4,
1647
+ "grad_norm": 16.80458641052246,
1648
+ "learning_rate": 1.5033333333333337e-06,
1649
+ "loss": 0.7020199298858643,
1650
+ "margin": 1.91923828125,
1651
+ "max_reward": 3.17734375,
1652
+ "mean_reward": 0.354052734375,
1653
+ "min_reward": -3.8123046875,
1654
+ "num_tokens": 1691904.0,
1655
+ "step": 1050
1656
+ },
1657
+ {
1658
+ "epoch": 1.4,
1659
+ "eval_accuracy": 0.606,
1660
+ "eval_loss": 1.3567456007003784,
1661
+ "eval_margin": 1.0848167724609374,
1662
+ "eval_max_reward": 1.356125732421875,
1663
+ "eval_mean_reward": -0.37665606689453124,
1664
+ "eval_min_reward": -2.1094378662109374,
1665
+ "eval_num_tokens": 1691904.0,
1666
+ "eval_runtime": 62.3936,
1667
+ "eval_samples_per_second": 8.014,
1668
+ "eval_steps_per_second": 8.014,
1669
+ "step": 1050
1670
+ },
1671
+ {
1672
+ "accuracy": 0.55,
1673
+ "epoch": 1.4133333333333333,
1674
+ "grad_norm": 117.69019317626953,
1675
+ "learning_rate": 1.4700000000000001e-06,
1676
+ "loss": 1.212506103515625,
1677
+ "margin": 1.3871475219726563,
1678
+ "max_reward": 2.3145721435546873,
1679
+ "mean_reward": -3.046747589111328,
1680
+ "min_reward": -6.746875,
1681
+ "num_tokens": 1708299.0,
1682
+ "step": 1060
1683
+ },
1684
+ {
1685
+ "accuracy": 0.6,
1686
+ "epoch": 1.4266666666666667,
1687
+ "grad_norm": 100.34335327148438,
1688
+ "learning_rate": 1.4366666666666667e-06,
1689
+ "loss": 0.9856803894042969,
1690
+ "margin": 1.3234375,
1691
+ "max_reward": 1.334375,
1692
+ "mean_reward": -2.6033203125,
1693
+ "min_reward": -6.49765625,
1694
+ "num_tokens": 1725007.0,
1695
+ "step": 1070
1696
+ },
1697
+ {
1698
+ "accuracy": 0.7,
1699
+ "epoch": 1.44,
1700
+ "grad_norm": 185.3729248046875,
1701
+ "learning_rate": 1.4033333333333336e-06,
1702
+ "loss": 1.1272445678710938,
1703
+ "margin": 2.0210784912109374,
1704
+ "max_reward": 3.303125,
1705
+ "mean_reward": -0.05469512939453125,
1706
+ "min_reward": -3.634649658203125,
1707
+ "num_tokens": 1740533.0,
1708
+ "step": 1080
1709
+ },
1710
+ {
1711
+ "accuracy": 0.6,
1712
+ "epoch": 1.4533333333333334,
1713
+ "grad_norm": 161.89877319335938,
1714
+ "learning_rate": 1.3700000000000002e-06,
1715
+ "loss": 1.4939114570617675,
1716
+ "margin": 0.5677734375,
1717
+ "max_reward": 1.749609375,
1718
+ "mean_reward": -2.25419921875,
1719
+ "min_reward": -6.180859375,
1720
+ "num_tokens": 1757099.0,
1721
+ "step": 1090
1722
+ },
1723
+ {
1724
+ "accuracy": 0.75,
1725
+ "epoch": 1.4666666666666668,
1726
+ "grad_norm": 119.40484619140625,
1727
+ "learning_rate": 1.3366666666666666e-06,
1728
+ "loss": 0.8826108932495117,
1729
+ "margin": 1.85439453125,
1730
+ "max_reward": 2.553125,
1731
+ "mean_reward": -0.852197265625,
1732
+ "min_reward": -4.6171875,
1733
+ "num_tokens": 1775146.0,
1734
+ "step": 1100
1735
+ },
1736
+ {
1737
+ "epoch": 1.4666666666666668,
1738
+ "eval_accuracy": 0.604,
1739
+ "eval_loss": 0.9978691339492798,
1740
+ "eval_margin": 0.7300045776367188,
1741
+ "eval_max_reward": 0.6507423095703125,
1742
+ "eval_mean_reward": -0.4976886901855469,
1743
+ "eval_min_reward": -1.6461196899414063,
1744
+ "eval_num_tokens": 1775146.0,
1745
+ "eval_runtime": 62.4188,
1746
+ "eval_samples_per_second": 8.01,
1747
+ "eval_steps_per_second": 8.01,
1748
+ "step": 1100
1749
+ },
1750
+ {
1751
+ "accuracy": 0.65,
1752
+ "epoch": 1.48,
1753
+ "grad_norm": 109.5771713256836,
1754
+ "learning_rate": 1.3033333333333335e-06,
1755
+ "loss": 0.9664765357971191,
1756
+ "margin": 0.737762451171875,
1757
+ "max_reward": 1.653662109375,
1758
+ "mean_reward": -0.5663665771484375,
1759
+ "min_reward": -3.6734375,
1760
+ "num_tokens": 1793063.0,
1761
+ "step": 1110
1762
+ },
1763
+ {
1764
+ "accuracy": 0.7,
1765
+ "epoch": 1.4933333333333334,
1766
+ "grad_norm": 83.6311264038086,
1767
+ "learning_rate": 1.2700000000000001e-06,
1768
+ "loss": 0.6062012672424316,
1769
+ "margin": 2.137274169921875,
1770
+ "max_reward": 1.9431640625,
1771
+ "mean_reward": -1.1080963134765625,
1772
+ "min_reward": -4.4921875,
1773
+ "num_tokens": 1808403.0,
1774
+ "step": 1120
1775
+ },
1776
+ {
1777
+ "accuracy": 0.75,
1778
+ "epoch": 1.5066666666666668,
1779
+ "grad_norm": 60.31748962402344,
1780
+ "learning_rate": 1.2366666666666668e-06,
1781
+ "loss": 0.7049060821533203,
1782
+ "margin": 1.48701171875,
1783
+ "max_reward": 2.08974609375,
1784
+ "mean_reward": -0.200439453125,
1785
+ "min_reward": -2.463671875,
1786
+ "num_tokens": 1824621.0,
1787
+ "step": 1130
1788
+ },
1789
+ {
1790
+ "accuracy": 0.55,
1791
+ "epoch": 1.52,
1792
+ "grad_norm": 149.60330200195312,
1793
+ "learning_rate": 1.2033333333333334e-06,
1794
+ "loss": 1.1337081909179687,
1795
+ "margin": 0.48056640625,
1796
+ "max_reward": 1.880859375,
1797
+ "mean_reward": -0.384716796875,
1798
+ "min_reward": -2.83515625,
1799
+ "num_tokens": 1840744.0,
1800
+ "step": 1140
1801
+ },
1802
+ {
1803
+ "accuracy": 0.7,
1804
+ "epoch": 1.5333333333333332,
1805
+ "grad_norm": 41.09004592895508,
1806
+ "learning_rate": 1.1700000000000002e-06,
1807
+ "loss": 0.6269871711730957,
1808
+ "margin": 1.2066650390625,
1809
+ "max_reward": 2.41328125,
1810
+ "mean_reward": 0.39190673828125,
1811
+ "min_reward": -1.7814453125,
1812
+ "num_tokens": 1855354.0,
1813
+ "step": 1150
1814
+ },
1815
+ {
1816
+ "epoch": 1.5333333333333332,
1817
+ "eval_accuracy": 0.586,
1818
+ "eval_loss": 0.9035583138465881,
1819
+ "eval_margin": 0.5397208251953125,
1820
+ "eval_max_reward": 1.6990680236816407,
1821
+ "eval_mean_reward": 0.7785431518554687,
1822
+ "eval_min_reward": -0.14198171997070314,
1823
+ "eval_num_tokens": 1855354.0,
1824
+ "eval_runtime": 62.4733,
1825
+ "eval_samples_per_second": 8.003,
1826
+ "eval_steps_per_second": 8.003,
1827
+ "step": 1150
1828
+ },
1829
+ {
1830
+ "accuracy": 0.85,
1831
+ "epoch": 1.5466666666666666,
1832
+ "grad_norm": 6.0896782875061035,
1833
+ "learning_rate": 1.1366666666666667e-06,
1834
+ "loss": 0.5843002796173096,
1835
+ "margin": 1.83350830078125,
1836
+ "max_reward": 2.8265625,
1837
+ "mean_reward": 0.397064208984375,
1838
+ "min_reward": -2.0440673828125,
1839
+ "num_tokens": 1871860.0,
1840
+ "step": 1160
1841
+ },
1842
+ {
1843
+ "accuracy": 0.75,
1844
+ "epoch": 1.56,
1845
+ "grad_norm": 102.23207092285156,
1846
+ "learning_rate": 1.1033333333333335e-06,
1847
+ "loss": 0.5763196468353271,
1848
+ "margin": 1.4167724609375,
1849
+ "max_reward": 3.4875,
1850
+ "mean_reward": 1.43270263671875,
1851
+ "min_reward": -1.278466796875,
1852
+ "num_tokens": 1888601.0,
1853
+ "step": 1170
1854
+ },
1855
+ {
1856
+ "accuracy": 0.6,
1857
+ "epoch": 1.5733333333333333,
1858
+ "grad_norm": 109.7595443725586,
1859
+ "learning_rate": 1.0700000000000001e-06,
1860
+ "loss": 0.9253439903259277,
1861
+ "margin": 0.2265625,
1862
+ "max_reward": 2.1390625,
1863
+ "mean_reward": 0.3791015625,
1864
+ "min_reward": -1.2005859375,
1865
+ "num_tokens": 1905756.0,
1866
+ "step": 1180
1867
+ },
1868
+ {
1869
+ "accuracy": 0.7,
1870
+ "epoch": 1.5866666666666667,
1871
+ "grad_norm": 27.03172492980957,
1872
+ "learning_rate": 1.0366666666666668e-06,
1873
+ "loss": 0.6028451919555664,
1874
+ "margin": 1.306298828125,
1875
+ "max_reward": 2.61015625,
1876
+ "mean_reward": 0.2985595703125,
1877
+ "min_reward": -1.97734375,
1878
+ "num_tokens": 1922025.0,
1879
+ "step": 1190
1880
+ },
1881
+ {
1882
+ "accuracy": 0.75,
1883
+ "epoch": 1.6,
1884
+ "grad_norm": 91.80132293701172,
1885
+ "learning_rate": 1.0033333333333334e-06,
1886
+ "loss": 0.6056195735931397,
1887
+ "margin": 1.401220703125,
1888
+ "max_reward": 2.41943359375,
1889
+ "mean_reward": -0.0202880859375,
1890
+ "min_reward": -2.3765625,
1891
+ "num_tokens": 1938931.0,
1892
+ "step": 1200
1893
+ },
1894
+ {
1895
+ "epoch": 1.6,
1896
+ "eval_accuracy": 0.602,
1897
+ "eval_loss": 1.04457688331604,
1898
+ "eval_margin": 0.740804443359375,
1899
+ "eval_max_reward": 1.606439208984375,
1900
+ "eval_mean_reward": 0.3948948974609375,
1901
+ "eval_min_reward": -0.8166494140625,
1902
+ "eval_num_tokens": 1938931.0,
1903
+ "eval_runtime": 62.3655,
1904
+ "eval_samples_per_second": 8.017,
1905
+ "eval_steps_per_second": 8.017,
1906
+ "step": 1200
1907
+ },
1908
+ {
1909
+ "accuracy": 0.75,
1910
+ "epoch": 1.6133333333333333,
1911
+ "grad_norm": 0.7046879529953003,
1912
+ "learning_rate": 9.7e-07,
1913
+ "loss": 1.0164477348327636,
1914
+ "margin": 1.867919921875,
1915
+ "max_reward": 3.298828125,
1916
+ "mean_reward": 0.1361083984375,
1917
+ "min_reward": -3.41015625,
1918
+ "num_tokens": 1955127.0,
1919
+ "step": 1210
1920
+ },
1921
+ {
1922
+ "accuracy": 0.65,
1923
+ "epoch": 1.6266666666666667,
1924
+ "grad_norm": 25.504459381103516,
1925
+ "learning_rate": 9.366666666666667e-07,
1926
+ "loss": 0.8065451622009278,
1927
+ "margin": 1.255029296875,
1928
+ "max_reward": 2.398828125,
1929
+ "mean_reward": -0.2284912109375,
1930
+ "min_reward": -3.13154296875,
1931
+ "num_tokens": 1972318.0,
1932
+ "step": 1220
1933
+ },
1934
+ {
1935
+ "accuracy": 0.65,
1936
+ "epoch": 1.6400000000000001,
1937
+ "grad_norm": 87.41260528564453,
1938
+ "learning_rate": 9.033333333333334e-07,
1939
+ "loss": 1.119236946105957,
1940
+ "margin": 0.7582611083984375,
1941
+ "max_reward": 2.4771484375,
1942
+ "mean_reward": -0.6033004760742188,
1943
+ "min_reward": -3.24765625,
1944
+ "num_tokens": 1987620.0,
1945
+ "step": 1230
1946
+ },
1947
+ {
1948
+ "accuracy": 0.65,
1949
+ "epoch": 1.6533333333333333,
1950
+ "grad_norm": 94.04547119140625,
1951
+ "learning_rate": 8.7e-07,
1952
+ "loss": 0.6439452171325684,
1953
+ "margin": 1.82333984375,
1954
+ "max_reward": 2.728125,
1955
+ "mean_reward": 0.210400390625,
1956
+ "min_reward": -2.7384765625,
1957
+ "num_tokens": 2004302.0,
1958
+ "step": 1240
1959
+ },
1960
+ {
1961
+ "accuracy": 0.65,
1962
+ "epoch": 1.6666666666666665,
1963
+ "grad_norm": 85.90032958984375,
1964
+ "learning_rate": 8.366666666666668e-07,
1965
+ "loss": 0.5457091331481934,
1966
+ "margin": 1.3715576171875,
1967
+ "max_reward": 2.209033203125,
1968
+ "mean_reward": -0.44652099609375,
1969
+ "min_reward": -2.5546875,
1970
+ "num_tokens": 2021262.0,
1971
+ "step": 1250
1972
+ },
1973
+ {
1974
+ "epoch": 1.6666666666666665,
1975
+ "eval_accuracy": 0.612,
1976
+ "eval_loss": 1.0178015232086182,
1977
+ "eval_margin": 0.7143900146484375,
1978
+ "eval_max_reward": 1.167669921875,
1979
+ "eval_mean_reward": 0.00498980712890625,
1980
+ "eval_min_reward": -1.1576903076171876,
1981
+ "eval_num_tokens": 2021262.0,
1982
+ "eval_runtime": 62.4994,
1983
+ "eval_samples_per_second": 8.0,
1984
+ "eval_steps_per_second": 8.0,
1985
+ "step": 1250
1986
+ },
1987
+ {
1988
+ "accuracy": 0.85,
1989
+ "epoch": 1.6800000000000002,
1990
+ "grad_norm": 57.4267578125,
1991
+ "learning_rate": 8.033333333333335e-07,
1992
+ "loss": 0.5381040573120117,
1993
+ "margin": 1.5513671875,
1994
+ "max_reward": 2.2685546875,
1995
+ "mean_reward": -0.218896484375,
1996
+ "min_reward": -2.88603515625,
1997
+ "num_tokens": 2037311.0,
1998
+ "step": 1260
1999
+ },
2000
+ {
2001
+ "accuracy": 0.65,
2002
+ "epoch": 1.6933333333333334,
2003
+ "grad_norm": 20.574970245361328,
2004
+ "learning_rate": 7.7e-07,
2005
+ "loss": 0.8994636535644531,
2006
+ "margin": 1.5009521484375,
2007
+ "max_reward": 2.207861328125,
2008
+ "mean_reward": -0.48370361328125,
2009
+ "min_reward": -2.67890625,
2010
+ "num_tokens": 2053587.0,
2011
+ "step": 1270
2012
+ },
2013
+ {
2014
+ "accuracy": 0.5,
2015
+ "epoch": 1.7066666666666666,
2016
+ "grad_norm": 25.6247501373291,
2017
+ "learning_rate": 7.366666666666668e-07,
2018
+ "loss": 1.4920889854431152,
2019
+ "margin": 0.5493316650390625,
2020
+ "max_reward": 2.29677734375,
2021
+ "mean_reward": -0.34183807373046876,
2022
+ "min_reward": -3.459832763671875,
2023
+ "num_tokens": 2069715.0,
2024
+ "step": 1280
2025
+ },
2026
+ {
2027
+ "accuracy": 0.65,
2028
+ "epoch": 1.72,
2029
+ "grad_norm": 28.003578186035156,
2030
+ "learning_rate": 7.033333333333334e-07,
2031
+ "loss": 1.0490900039672852,
2032
+ "margin": 1.029248046875,
2033
+ "max_reward": 2.55390625,
2034
+ "mean_reward": -0.0261474609375,
2035
+ "min_reward": -2.5232421875,
2036
+ "num_tokens": 2083729.0,
2037
+ "step": 1290
2038
+ },
2039
+ {
2040
+ "accuracy": 0.6,
2041
+ "epoch": 1.7333333333333334,
2042
+ "grad_norm": 100.13626098632812,
2043
+ "learning_rate": 6.7e-07,
2044
+ "loss": 0.8861888885498047,
2045
+ "margin": 0.58553466796875,
2046
+ "max_reward": 1.53046875,
2047
+ "mean_reward": -0.445611572265625,
2048
+ "min_reward": -2.4734375,
2049
+ "num_tokens": 2099948.0,
2050
+ "step": 1300
2051
+ },
2052
+ {
2053
+ "epoch": 1.7333333333333334,
2054
+ "eval_accuracy": 0.606,
2055
+ "eval_loss": 1.0033907890319824,
2056
+ "eval_margin": 0.7165312805175781,
2057
+ "eval_max_reward": 1.2282371826171874,
2058
+ "eval_mean_reward": 0.07899192810058593,
2059
+ "eval_min_reward": -1.0702533264160157,
2060
+ "eval_num_tokens": 2099948.0,
2061
+ "eval_runtime": 62.3657,
2062
+ "eval_samples_per_second": 8.017,
2063
+ "eval_steps_per_second": 8.017,
2064
+ "step": 1300
2065
+ },
2066
+ {
2067
+ "accuracy": 0.7,
2068
+ "epoch": 1.7466666666666666,
2069
+ "grad_norm": 144.0567626953125,
2070
+ "learning_rate": 6.366666666666667e-07,
2071
+ "loss": 0.5809046268463135,
2072
+ "margin": 1.460107421875,
2073
+ "max_reward": 1.62353515625,
2074
+ "mean_reward": -0.3951416015625,
2075
+ "min_reward": -2.947265625,
2076
+ "num_tokens": 2115789.0,
2077
+ "step": 1310
2078
+ },
2079
+ {
2080
+ "accuracy": 0.7,
2081
+ "epoch": 1.76,
2082
+ "grad_norm": 49.05362319946289,
2083
+ "learning_rate": 6.033333333333334e-07,
2084
+ "loss": 0.87122802734375,
2085
+ "margin": 1.10679931640625,
2086
+ "max_reward": 2.268359375,
2087
+ "mean_reward": -0.050628662109375,
2088
+ "min_reward": -2.1308349609375,
2089
+ "num_tokens": 2133528.0,
2090
+ "step": 1320
2091
+ },
2092
+ {
2093
+ "accuracy": 0.65,
2094
+ "epoch": 1.7733333333333334,
2095
+ "grad_norm": 98.22369384765625,
2096
+ "learning_rate": 5.7e-07,
2097
+ "loss": 0.6749707221984863,
2098
+ "margin": 0.971484375,
2099
+ "max_reward": 2.073046875,
2100
+ "mean_reward": -0.6001953125,
2101
+ "min_reward": -3.078125,
2102
+ "num_tokens": 2147275.0,
2103
+ "step": 1330
2104
+ },
2105
+ {
2106
+ "accuracy": 0.75,
2107
+ "epoch": 1.7866666666666666,
2108
+ "grad_norm": 178.06886291503906,
2109
+ "learning_rate": 5.366666666666667e-07,
2110
+ "loss": 0.7042564868927002,
2111
+ "margin": 1.1953125,
2112
+ "max_reward": 1.126953125,
2113
+ "mean_reward": -1.3919921875,
2114
+ "min_reward": -3.98046875,
2115
+ "num_tokens": 2165353.0,
2116
+ "step": 1340
2117
+ },
2118
+ {
2119
+ "accuracy": 0.65,
2120
+ "epoch": 1.8,
2121
+ "grad_norm": 49.11288070678711,
2122
+ "learning_rate": 5.033333333333334e-07,
2123
+ "loss": 0.9773781776428223,
2124
+ "margin": 0.57353515625,
2125
+ "max_reward": 1.633447265625,
2126
+ "mean_reward": -0.50439453125,
2127
+ "min_reward": -2.768359375,
2128
+ "num_tokens": 2181126.0,
2129
+ "step": 1350
2130
+ },
2131
+ {
2132
+ "epoch": 1.8,
2133
+ "eval_accuracy": 0.612,
2134
+ "eval_loss": 0.9391070008277893,
2135
+ "eval_margin": 0.6821452369689941,
2136
+ "eval_max_reward": 0.2567112731933594,
2137
+ "eval_mean_reward": -0.8024592456817627,
2138
+ "eval_min_reward": -1.8616297645568847,
2139
+ "eval_num_tokens": 2181126.0,
2140
+ "eval_runtime": 62.4622,
2141
+ "eval_samples_per_second": 8.005,
2142
+ "eval_steps_per_second": 8.005,
2143
+ "step": 1350
2144
+ },
2145
+ {
2146
+ "accuracy": 0.8,
2147
+ "epoch": 1.8133333333333335,
2148
+ "grad_norm": 12.369355201721191,
2149
+ "learning_rate": 4.7000000000000005e-07,
2150
+ "loss": 0.5082510471343994,
2151
+ "margin": 2.278076171875,
2152
+ "max_reward": 2.44150390625,
2153
+ "mean_reward": -0.1749755859375,
2154
+ "min_reward": -3.0359375,
2155
+ "num_tokens": 2198470.0,
2156
+ "step": 1360
2157
+ },
2158
+ {
2159
+ "accuracy": 0.6,
2160
+ "epoch": 1.8266666666666667,
2161
+ "grad_norm": 21.943758010864258,
2162
+ "learning_rate": 4.366666666666667e-07,
2163
+ "loss": 0.8553391456604004,
2164
+ "margin": 1.10614013671875,
2165
+ "max_reward": 0.610546875,
2166
+ "mean_reward": -1.453851318359375,
2167
+ "min_reward": -3.56328125,
2168
+ "num_tokens": 2213797.0,
2169
+ "step": 1370
2170
+ },
2171
+ {
2172
+ "accuracy": 0.55,
2173
+ "epoch": 1.8399999999999999,
2174
+ "grad_norm": 67.34660339355469,
2175
+ "learning_rate": 4.0333333333333337e-07,
2176
+ "loss": 1.136549472808838,
2177
+ "margin": 0.629931640625,
2178
+ "max_reward": 2.42353515625,
2179
+ "mean_reward": -0.6287841796875,
2180
+ "min_reward": -3.7546875,
2181
+ "num_tokens": 2230299.0,
2182
+ "step": 1380
2183
+ },
2184
+ {
2185
+ "accuracy": 0.65,
2186
+ "epoch": 1.8533333333333335,
2187
+ "grad_norm": 14.513836860656738,
2188
+ "learning_rate": 3.7e-07,
2189
+ "loss": 0.5644237518310546,
2190
+ "margin": 1.4962158203125,
2191
+ "max_reward": 1.477099609375,
2192
+ "mean_reward": -1.04271240234375,
2193
+ "min_reward": -3.7123046875,
2194
+ "num_tokens": 2245618.0,
2195
+ "step": 1390
2196
+ },
2197
+ {
2198
+ "accuracy": 0.5,
2199
+ "epoch": 1.8666666666666667,
2200
+ "grad_norm": 137.32266235351562,
2201
+ "learning_rate": 3.366666666666667e-07,
2202
+ "loss": 0.7403782367706299,
2203
+ "margin": 1.222882080078125,
2204
+ "max_reward": 1.35396728515625,
2205
+ "mean_reward": -1.0289886474609375,
2206
+ "min_reward": -3.173046875,
2207
+ "num_tokens": 2262798.0,
2208
+ "step": 1400
2209
+ },
2210
+ {
2211
+ "epoch": 1.8666666666666667,
2212
+ "eval_accuracy": 0.618,
2213
+ "eval_loss": 0.9525383710861206,
2214
+ "eval_margin": 0.7151371459960938,
2215
+ "eval_max_reward": -0.032703857421875,
2216
+ "eval_mean_reward": -1.126583221435547,
2217
+ "eval_min_reward": -2.220462585449219,
2218
+ "eval_num_tokens": 2262798.0,
2219
+ "eval_runtime": 62.453,
2220
+ "eval_samples_per_second": 8.006,
2221
+ "eval_steps_per_second": 8.006,
2222
+ "step": 1400
2223
+ },
2224
+ {
2225
+ "accuracy": 0.75,
2226
+ "epoch": 1.88,
2227
+ "grad_norm": 87.93866729736328,
2228
+ "learning_rate": 3.033333333333334e-07,
2229
+ "loss": 0.7155962944030761,
2230
+ "margin": 1.211572265625,
2231
+ "max_reward": 1.64248046875,
2232
+ "mean_reward": -0.8534912109375,
2233
+ "min_reward": -3.3265625,
2234
+ "num_tokens": 2280625.0,
2235
+ "step": 1410
2236
+ },
2237
+ {
2238
+ "accuracy": 0.6,
2239
+ "epoch": 1.8933333333333333,
2240
+ "grad_norm": 136.62033081054688,
2241
+ "learning_rate": 2.7e-07,
2242
+ "loss": 0.7803818702697753,
2243
+ "margin": 1.5741943359375,
2244
+ "max_reward": 1.125,
2245
+ "mean_reward": -1.46397705078125,
2246
+ "min_reward": -3.69921875,
2247
+ "num_tokens": 2297438.0,
2248
+ "step": 1420
2249
+ },
2250
+ {
2251
+ "accuracy": 0.8,
2252
+ "epoch": 1.9066666666666667,
2253
+ "grad_norm": 36.08385467529297,
2254
+ "learning_rate": 2.3666666666666667e-07,
2255
+ "loss": 0.44714956283569335,
2256
+ "margin": 1.790771484375,
2257
+ "max_reward": 2.1365234375,
2258
+ "mean_reward": -1.1315673828125,
2259
+ "min_reward": -4.0796875,
2260
+ "num_tokens": 2315559.0,
2261
+ "step": 1430
2262
+ },
2263
+ {
2264
+ "accuracy": 0.75,
2265
+ "epoch": 1.92,
2266
+ "grad_norm": 16.93815803527832,
2267
+ "learning_rate": 2.0333333333333333e-07,
2268
+ "loss": 0.6091221332550049,
2269
+ "margin": 1.25640869140625,
2270
+ "max_reward": 0.8255126953125,
2271
+ "mean_reward": -1.083807373046875,
2272
+ "min_reward": -3.1236328125,
2273
+ "num_tokens": 2331603.0,
2274
+ "step": 1440
2275
+ },
2276
+ {
2277
+ "accuracy": 0.55,
2278
+ "epoch": 1.9333333333333333,
2279
+ "grad_norm": 53.92049026489258,
2280
+ "learning_rate": 1.7000000000000001e-07,
2281
+ "loss": 0.7587657928466797,
2282
+ "margin": 1.339501953125,
2283
+ "max_reward": 1.06416015625,
2284
+ "mean_reward": -0.8656494140625,
2285
+ "min_reward": -3.60859375,
2286
+ "num_tokens": 2346680.0,
2287
+ "step": 1450
2288
+ },
2289
+ {
2290
+ "epoch": 1.9333333333333333,
2291
+ "eval_accuracy": 0.612,
2292
+ "eval_loss": 0.97000652551651,
2293
+ "eval_margin": 0.7449773254394532,
2294
+ "eval_max_reward": 0.27565774536132814,
2295
+ "eval_mean_reward": -0.8554528350830078,
2296
+ "eval_min_reward": -1.9865634155273437,
2297
+ "eval_num_tokens": 2346680.0,
2298
+ "eval_runtime": 62.49,
2299
+ "eval_samples_per_second": 8.001,
2300
+ "eval_steps_per_second": 8.001,
2301
+ "step": 1450
2302
+ }
2303
+ ],
2304
+ "logging_steps": 10,
2305
+ "max_steps": 1500,
2306
+ "num_input_tokens_seen": 0,
2307
+ "num_train_epochs": 2,
2308
+ "save_steps": 50,
2309
+ "stateful_callbacks": {
2310
+ "TrainerControl": {
2311
+ "args": {
2312
+ "should_epoch_stop": false,
2313
+ "should_evaluate": false,
2314
+ "should_log": false,
2315
+ "should_save": true,
2316
+ "should_training_stop": false
2317
+ },
2318
+ "attributes": {}
2319
+ }
2320
+ },
2321
+ "total_flos": 1987610824820736.0,
2322
+ "train_batch_size": 2,
2323
+ "trial_name": null,
2324
+ "trial_params": null
2325
+ }
checkpoint-1450/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d5bcd49bd6d1025adc5d4385da3c333196b4008baf24052fee3d16f53db65152
3
+ size 5393
checkpoint-1500/config.json ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "activation_function": "gelu_pytorch_tanh",
3
+ "add_cross_attention": false,
4
+ "architectures": [
5
+ "GPTBigCodeForSequenceClassification"
6
+ ],
7
+ "attention_softmax_in_fp32": true,
8
+ "attn_pdrop": 0.1,
9
+ "bos_token_id": 0,
10
+ "dtype": "float32",
11
+ "embd_pdrop": 0.1,
12
+ "eos_token_id": 0,
13
+ "id2label": {
14
+ "0": "LABEL_0"
15
+ },
16
+ "inference_runner": 0,
17
+ "initializer_range": 0.02,
18
+ "label2id": {
19
+ "LABEL_0": 0
20
+ },
21
+ "layer_norm_epsilon": 1e-05,
22
+ "max_batch_size": null,
23
+ "max_sequence_length": null,
24
+ "model_type": "gpt_bigcode",
25
+ "multi_query": true,
26
+ "n_embd": 768,
27
+ "n_head": 12,
28
+ "n_inner": 3072,
29
+ "n_layer": 20,
30
+ "n_positions": 8192,
31
+ "num_key_value_heads": 1,
32
+ "pad_key_length": true,
33
+ "pad_token_id": 0,
34
+ "pre_allocate_kv_cache": false,
35
+ "resid_pdrop": 0.1,
36
+ "scale_attention_softmax_in_fp32": true,
37
+ "scale_attn_weights": true,
38
+ "summary_activation": null,
39
+ "summary_first_dropout": 0.1,
40
+ "summary_proj_to_labels": true,
41
+ "summary_type": "cls_index",
42
+ "summary_use_proj": true,
43
+ "tie_word_embeddings": true,
44
+ "transformers_version": "5.0.0",
45
+ "use_cache": false,
46
+ "validate_runner_input": true,
47
+ "vocab_size": 49152
48
+ }
checkpoint-1500/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:700a6584a30ad5bbfe191ea3c16f9cc5f1c1f94c65d782fa167b907d55fbe920
3
+ size 656604448
checkpoint-1500/optimizer.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d4e6c0be9cd9034b38d78d07ecd76ad0b0cf231abad02f4803afbd63c7e597c6
3
+ size 1313367755
checkpoint-1500/rng_state.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7ec9e4379c2bb6bfe23011594c8868fe099b0f7a368c6a17618411287f59c9f3
3
+ size 14645
checkpoint-1500/scheduler.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:df84d2bf98e4531bd9b6bb1a8a22f57b8c890c451f81a06449b06b2ae7d0af93
3
+ size 1465
checkpoint-1500/tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
checkpoint-1500/tokenizer_config.json ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "backend": "tokenizers",
4
+ "bos_token": "<|endoftext|>",
5
+ "eos_token": "<|endoftext|>",
6
+ "errors": "replace",
7
+ "extra_special_tokens": [
8
+ "<|endoftext|>",
9
+ "<fim_prefix>",
10
+ "<fim_middle>",
11
+ "<fim_suffix>",
12
+ "<fim_pad>",
13
+ "<filename>",
14
+ "<gh_stars>",
15
+ "<issue_start>",
16
+ "<issue_comment>",
17
+ "<issue_closed>",
18
+ "<jupyter_start>",
19
+ "<jupyter_text>",
20
+ "<jupyter_code>",
21
+ "<jupyter_output>",
22
+ "<empty_output>",
23
+ "<commit_before>",
24
+ "<commit_msg>",
25
+ "<commit_after>",
26
+ "<reponame>"
27
+ ],
28
+ "is_local": false,
29
+ "model_max_length": 1000000000000000019884624838656,
30
+ "pad_token": "<|endoftext|>",
31
+ "tokenizer_class": "GPT2Tokenizer",
32
+ "unk_token": "<|endoftext|>",
33
+ "vocab_size": 49152
34
+ }
checkpoint-1500/trainer_state.json ADDED
@@ -0,0 +1,2404 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_global_step": null,
3
+ "best_metric": null,
4
+ "best_model_checkpoint": null,
5
+ "epoch": 2.0,
6
+ "eval_steps": 50,
7
+ "global_step": 1500,
8
+ "is_hyper_param_search": false,
9
+ "is_local_process_zero": true,
10
+ "is_world_process_zero": true,
11
+ "log_history": [
12
+ {
13
+ "accuracy": 0.65,
14
+ "epoch": 0.013333333333333334,
15
+ "grad_norm": 26.781282424926758,
16
+ "learning_rate": 4.970000000000001e-06,
17
+ "loss": 0.6315235137939453,
18
+ "margin": 0.18203125,
19
+ "max_reward": -1.005078125,
20
+ "mean_reward": -1.4798828125,
21
+ "min_reward": -1.972265625,
22
+ "num_tokens": 16214.0,
23
+ "step": 10
24
+ },
25
+ {
26
+ "accuracy": 0.35,
27
+ "epoch": 0.02666666666666667,
28
+ "grad_norm": 18.78548812866211,
29
+ "learning_rate": 4.936666666666667e-06,
30
+ "loss": 0.9058952331542969,
31
+ "margin": -0.21234130859375,
32
+ "max_reward": -0.0454345703125,
33
+ "mean_reward": -1.138067626953125,
34
+ "min_reward": -2.29296875,
35
+ "num_tokens": 32750.0,
36
+ "step": 20
37
+ },
38
+ {
39
+ "accuracy": 0.65,
40
+ "epoch": 0.04,
41
+ "grad_norm": 23.18710708618164,
42
+ "learning_rate": 4.9033333333333335e-06,
43
+ "loss": 0.666615343093872,
44
+ "margin": 0.1455078125,
45
+ "max_reward": -0.2720703125,
46
+ "mean_reward": -0.9361328125,
47
+ "min_reward": -1.67734375,
48
+ "num_tokens": 49136.0,
49
+ "step": 30
50
+ },
51
+ {
52
+ "accuracy": 0.55,
53
+ "epoch": 0.05333333333333334,
54
+ "grad_norm": 23.498512268066406,
55
+ "learning_rate": 4.87e-06,
56
+ "loss": 0.775050687789917,
57
+ "margin": -0.038720703125,
58
+ "max_reward": -0.69931640625,
59
+ "mean_reward": -1.3807861328125,
60
+ "min_reward": -1.86953125,
61
+ "num_tokens": 64999.0,
62
+ "step": 40
63
+ },
64
+ {
65
+ "accuracy": 0.7,
66
+ "epoch": 0.06666666666666667,
67
+ "grad_norm": 25.616416931152344,
68
+ "learning_rate": 4.836666666666667e-06,
69
+ "loss": 0.6898736953735352,
70
+ "margin": 0.26455078125,
71
+ "max_reward": -0.5318359375,
72
+ "mean_reward": -1.313720703125,
73
+ "min_reward": -2.13203125,
74
+ "num_tokens": 82563.0,
75
+ "step": 50
76
+ },
77
+ {
78
+ "epoch": 0.06666666666666667,
79
+ "eval_accuracy": 0.532,
80
+ "eval_loss": 0.7488015294075012,
81
+ "eval_margin": 0.0769696044921875,
82
+ "eval_max_reward": -1.1200030517578126,
83
+ "eval_mean_reward": -1.4668482055664063,
84
+ "eval_min_reward": -1.813693359375,
85
+ "eval_num_tokens": 82563.0,
86
+ "eval_runtime": 59.7497,
87
+ "eval_samples_per_second": 8.368,
88
+ "eval_steps_per_second": 8.368,
89
+ "step": 50
90
+ },
91
+ {
92
+ "accuracy": 0.5,
93
+ "epoch": 0.08,
94
+ "grad_norm": 47.123233795166016,
95
+ "learning_rate": 4.803333333333334e-06,
96
+ "loss": 0.7043657302856445,
97
+ "margin": 0.144775390625,
98
+ "max_reward": -0.82724609375,
99
+ "mean_reward": -1.5813232421875,
100
+ "min_reward": -2.2578125,
101
+ "num_tokens": 100158.0,
102
+ "step": 60
103
+ },
104
+ {
105
+ "accuracy": 0.45,
106
+ "epoch": 0.09333333333333334,
107
+ "grad_norm": 43.35136032104492,
108
+ "learning_rate": 4.77e-06,
109
+ "loss": 0.7232241630554199,
110
+ "margin": 0.070654296875,
111
+ "max_reward": -1.10107421875,
112
+ "mean_reward": -1.6929443359375,
113
+ "min_reward": -2.36640625,
114
+ "num_tokens": 116349.0,
115
+ "step": 70
116
+ },
117
+ {
118
+ "accuracy": 0.35,
119
+ "epoch": 0.10666666666666667,
120
+ "grad_norm": 38.70017623901367,
121
+ "learning_rate": 4.736666666666667e-06,
122
+ "loss": 0.9364012718200684,
123
+ "margin": -0.284765625,
124
+ "max_reward": -0.4794921875,
125
+ "mean_reward": -1.26298828125,
126
+ "min_reward": -1.909375,
127
+ "num_tokens": 132123.0,
128
+ "step": 80
129
+ },
130
+ {
131
+ "accuracy": 0.45,
132
+ "epoch": 0.12,
133
+ "grad_norm": 29.024276733398438,
134
+ "learning_rate": 4.703333333333334e-06,
135
+ "loss": 0.8137557983398438,
136
+ "margin": -0.1231597900390625,
137
+ "max_reward": 0.445391845703125,
138
+ "mean_reward": -0.22428436279296876,
139
+ "min_reward": -0.788671875,
140
+ "num_tokens": 147982.0,
141
+ "step": 90
142
+ },
143
+ {
144
+ "accuracy": 0.45,
145
+ "epoch": 0.13333333333333333,
146
+ "grad_norm": 34.12649154663086,
147
+ "learning_rate": 4.670000000000001e-06,
148
+ "loss": 0.7763675689697266,
149
+ "margin": -0.06761474609375,
150
+ "max_reward": 0.74072265625,
151
+ "mean_reward": 0.255206298828125,
152
+ "min_reward": -0.2246826171875,
153
+ "num_tokens": 164058.0,
154
+ "step": 100
155
+ },
156
+ {
157
+ "epoch": 0.13333333333333333,
158
+ "eval_accuracy": 0.51,
159
+ "eval_loss": 0.7375237941741943,
160
+ "eval_margin": 0.008534343719482421,
161
+ "eval_max_reward": 0.5244478607177734,
162
+ "eval_mean_reward": 0.2890134220123291,
163
+ "eval_min_reward": 0.053578983306884764,
164
+ "eval_num_tokens": 164058.0,
165
+ "eval_runtime": 62.906,
166
+ "eval_samples_per_second": 7.948,
167
+ "eval_steps_per_second": 7.948,
168
+ "step": 100
169
+ },
170
+ {
171
+ "accuracy": 0.45,
172
+ "epoch": 0.14666666666666667,
173
+ "grad_norm": 41.91898727416992,
174
+ "learning_rate": 4.6366666666666674e-06,
175
+ "loss": 0.7357148647308349,
176
+ "margin": -0.05003662109375,
177
+ "max_reward": 0.8333984375,
178
+ "mean_reward": 0.398614501953125,
179
+ "min_reward": -0.0110595703125,
180
+ "num_tokens": 180287.0,
181
+ "step": 110
182
+ },
183
+ {
184
+ "accuracy": 0.45,
185
+ "epoch": 0.16,
186
+ "grad_norm": 27.607086181640625,
187
+ "learning_rate": 4.603333333333334e-06,
188
+ "loss": 0.7425424575805664,
189
+ "margin": -0.05296630859375,
190
+ "max_reward": 0.73359375,
191
+ "mean_reward": 0.207928466796875,
192
+ "min_reward": -0.285498046875,
193
+ "num_tokens": 195395.0,
194
+ "step": 120
195
+ },
196
+ {
197
+ "accuracy": 0.55,
198
+ "epoch": 0.17333333333333334,
199
+ "grad_norm": 18.465009689331055,
200
+ "learning_rate": 4.57e-06,
201
+ "loss": 0.7002963066101074,
202
+ "margin": 0.0166015625,
203
+ "max_reward": 0.15341796875,
204
+ "mean_reward": -0.245654296875,
205
+ "min_reward": -0.6259765625,
206
+ "num_tokens": 211423.0,
207
+ "step": 130
208
+ },
209
+ {
210
+ "accuracy": 0.85,
211
+ "epoch": 0.18666666666666668,
212
+ "grad_norm": 24.311744689941406,
213
+ "learning_rate": 4.536666666666667e-06,
214
+ "loss": 0.5668483257293702,
215
+ "margin": 0.39488706588745115,
216
+ "max_reward": 0.4626953125,
217
+ "mean_reward": -0.23407502174377443,
218
+ "min_reward": -0.8857421875,
219
+ "num_tokens": 227155.0,
220
+ "step": 140
221
+ },
222
+ {
223
+ "accuracy": 0.55,
224
+ "epoch": 0.2,
225
+ "grad_norm": 27.90814971923828,
226
+ "learning_rate": 4.503333333333333e-06,
227
+ "loss": 0.9236076354980469,
228
+ "margin": 0.2251708984375,
229
+ "max_reward": 1.816796875,
230
+ "mean_reward": 0.43135986328125,
231
+ "min_reward": -1.107373046875,
232
+ "num_tokens": 243609.0,
233
+ "step": 150
234
+ },
235
+ {
236
+ "epoch": 0.2,
237
+ "eval_accuracy": 0.568,
238
+ "eval_loss": 0.9139922857284546,
239
+ "eval_margin": 0.21751553344726562,
240
+ "eval_max_reward": 1.1517796630859376,
241
+ "eval_mean_reward": 0.4644862823486328,
242
+ "eval_min_reward": -0.22280709838867188,
243
+ "eval_num_tokens": 243609.0,
244
+ "eval_runtime": 62.3021,
245
+ "eval_samples_per_second": 8.025,
246
+ "eval_steps_per_second": 8.025,
247
+ "step": 150
248
+ },
249
+ {
250
+ "accuracy": 0.5,
251
+ "epoch": 0.21333333333333335,
252
+ "grad_norm": 47.18056869506836,
253
+ "learning_rate": 4.47e-06,
254
+ "loss": 1.0227947235107422,
255
+ "margin": -0.06168212890625,
256
+ "max_reward": 1.634716796875,
257
+ "mean_reward": 0.546588134765625,
258
+ "min_reward": -0.413671875,
259
+ "num_tokens": 261657.0,
260
+ "step": 160
261
+ },
262
+ {
263
+ "accuracy": 0.5,
264
+ "epoch": 0.22666666666666666,
265
+ "grad_norm": 19.242305755615234,
266
+ "learning_rate": 4.436666666666667e-06,
267
+ "loss": 0.6836428642272949,
268
+ "margin": 0.1833953857421875,
269
+ "max_reward": 0.95390625,
270
+ "mean_reward": 0.25375518798828123,
271
+ "min_reward": -0.5401123046875,
272
+ "num_tokens": 277075.0,
273
+ "step": 170
274
+ },
275
+ {
276
+ "accuracy": 0.65,
277
+ "epoch": 0.24,
278
+ "grad_norm": 13.613568305969238,
279
+ "learning_rate": 4.403333333333334e-06,
280
+ "loss": 0.5862503528594971,
281
+ "margin": 0.335546875,
282
+ "max_reward": 0.95908203125,
283
+ "mean_reward": 0.436962890625,
284
+ "min_reward": -0.14921875,
285
+ "num_tokens": 291771.0,
286
+ "step": 180
287
+ },
288
+ {
289
+ "accuracy": 0.75,
290
+ "epoch": 0.25333333333333335,
291
+ "grad_norm": 16.98612403869629,
292
+ "learning_rate": 4.3700000000000005e-06,
293
+ "loss": 0.4962740898132324,
294
+ "margin": 0.72353515625,
295
+ "max_reward": 2.7734375,
296
+ "mean_reward": 1.643115234375,
297
+ "min_reward": 0.7041015625,
298
+ "num_tokens": 306869.0,
299
+ "step": 190
300
+ },
301
+ {
302
+ "accuracy": 0.55,
303
+ "epoch": 0.26666666666666666,
304
+ "grad_norm": 5.9575724601745605,
305
+ "learning_rate": 4.336666666666667e-06,
306
+ "loss": 0.8757614135742188,
307
+ "margin": 0.180908203125,
308
+ "max_reward": 3.7546875,
309
+ "mean_reward": 2.4668212890625,
310
+ "min_reward": 1.08994140625,
311
+ "num_tokens": 322209.0,
312
+ "step": 200
313
+ },
314
+ {
315
+ "epoch": 0.26666666666666666,
316
+ "eval_accuracy": 0.558,
317
+ "eval_loss": 0.8031588196754456,
318
+ "eval_margin": 0.211654052734375,
319
+ "eval_max_reward": 3.3186171875,
320
+ "eval_mean_reward": 2.7612215576171875,
321
+ "eval_min_reward": 2.203825927734375,
322
+ "eval_num_tokens": 322209.0,
323
+ "eval_runtime": 62.4301,
324
+ "eval_samples_per_second": 8.009,
325
+ "eval_steps_per_second": 8.009,
326
+ "step": 200
327
+ },
328
+ {
329
+ "accuracy": 0.5,
330
+ "epoch": 0.28,
331
+ "grad_norm": 20.00104331970215,
332
+ "learning_rate": 4.303333333333334e-06,
333
+ "loss": 0.742474889755249,
334
+ "margin": 0.2575439453125,
335
+ "max_reward": 3.06640625,
336
+ "mean_reward": 1.81966552734375,
337
+ "min_reward": 0.197021484375,
338
+ "num_tokens": 338225.0,
339
+ "step": 210
340
+ },
341
+ {
342
+ "accuracy": 0.5,
343
+ "epoch": 0.29333333333333333,
344
+ "grad_norm": 23.415834426879883,
345
+ "learning_rate": 4.270000000000001e-06,
346
+ "loss": 0.8180597305297852,
347
+ "margin": 0.2842529296875,
348
+ "max_reward": 4.1375,
349
+ "mean_reward": 2.89720458984375,
350
+ "min_reward": 1.491162109375,
351
+ "num_tokens": 352992.0,
352
+ "step": 220
353
+ },
354
+ {
355
+ "accuracy": 0.5,
356
+ "epoch": 0.30666666666666664,
357
+ "grad_norm": 28.938772201538086,
358
+ "learning_rate": 4.236666666666667e-06,
359
+ "loss": 0.8096456527709961,
360
+ "margin": -0.066015625,
361
+ "max_reward": 3.71875,
362
+ "mean_reward": 2.8822265625,
363
+ "min_reward": 1.9859375,
364
+ "num_tokens": 368570.0,
365
+ "step": 230
366
+ },
367
+ {
368
+ "accuracy": 0.65,
369
+ "epoch": 0.32,
370
+ "grad_norm": 19.143875122070312,
371
+ "learning_rate": 4.2033333333333335e-06,
372
+ "loss": 0.5784501552581787,
373
+ "margin": 0.43447265625,
374
+ "max_reward": 2.54921875,
375
+ "mean_reward": 1.823486328125,
376
+ "min_reward": 1.0931640625,
377
+ "num_tokens": 384794.0,
378
+ "step": 240
379
+ },
380
+ {
381
+ "accuracy": 0.6,
382
+ "epoch": 0.3333333333333333,
383
+ "grad_norm": 25.47344207763672,
384
+ "learning_rate": 4.17e-06,
385
+ "loss": 0.7684556484222412,
386
+ "margin": 0.04365234375,
387
+ "max_reward": 3.128125,
388
+ "mean_reward": 2.072509765625,
389
+ "min_reward": 1.1900390625,
390
+ "num_tokens": 401148.0,
391
+ "step": 250
392
+ },
393
+ {
394
+ "epoch": 0.3333333333333333,
395
+ "eval_accuracy": 0.572,
396
+ "eval_loss": 0.713194727897644,
397
+ "eval_margin": 0.1503466796875,
398
+ "eval_max_reward": 2.57871484375,
399
+ "eval_mean_reward": 2.22652490234375,
400
+ "eval_min_reward": 1.8743349609375,
401
+ "eval_num_tokens": 401148.0,
402
+ "eval_runtime": 62.3122,
403
+ "eval_samples_per_second": 8.024,
404
+ "eval_steps_per_second": 8.024,
405
+ "step": 250
406
+ },
407
+ {
408
+ "accuracy": 0.3,
409
+ "epoch": 0.3466666666666667,
410
+ "grad_norm": 27.635066986083984,
411
+ "learning_rate": 4.136666666666667e-06,
412
+ "loss": 0.9010297775268554,
413
+ "margin": -0.2201416015625,
414
+ "max_reward": 2.42265625,
415
+ "mean_reward": 1.72039794921875,
416
+ "min_reward": 0.801123046875,
417
+ "num_tokens": 416243.0,
418
+ "step": 260
419
+ },
420
+ {
421
+ "accuracy": 0.65,
422
+ "epoch": 0.36,
423
+ "grad_norm": 22.39845848083496,
424
+ "learning_rate": 4.1033333333333336e-06,
425
+ "loss": 0.7369675159454345,
426
+ "margin": 0.000390625,
427
+ "max_reward": 2.31796875,
428
+ "mean_reward": 1.5876953125,
429
+ "min_reward": 0.834375,
430
+ "num_tokens": 432425.0,
431
+ "step": 270
432
+ },
433
+ {
434
+ "accuracy": 0.6,
435
+ "epoch": 0.37333333333333335,
436
+ "grad_norm": 30.26738166809082,
437
+ "learning_rate": 4.07e-06,
438
+ "loss": 0.7203866004943847,
439
+ "margin": 0.00186767578125,
440
+ "max_reward": 1.509375,
441
+ "mean_reward": 1.016265869140625,
442
+ "min_reward": 0.5093994140625,
443
+ "num_tokens": 446709.0,
444
+ "step": 280
445
+ },
446
+ {
447
+ "accuracy": 0.6,
448
+ "epoch": 0.38666666666666666,
449
+ "grad_norm": 17.360742568969727,
450
+ "learning_rate": 4.036666666666667e-06,
451
+ "loss": 0.6374303817749023,
452
+ "margin": 0.159130859375,
453
+ "max_reward": 1.0638671875,
454
+ "mean_reward": 0.6469970703125,
455
+ "min_reward": 0.1876953125,
456
+ "num_tokens": 462422.0,
457
+ "step": 290
458
+ },
459
+ {
460
+ "accuracy": 0.45,
461
+ "epoch": 0.4,
462
+ "grad_norm": 24.10837173461914,
463
+ "learning_rate": 4.003333333333334e-06,
464
+ "loss": 0.7569143772125244,
465
+ "margin": 0.0091796875,
466
+ "max_reward": 1.9375,
467
+ "mean_reward": 1.299609375,
468
+ "min_reward": 0.747265625,
469
+ "num_tokens": 480306.0,
470
+ "step": 300
471
+ },
472
+ {
473
+ "epoch": 0.4,
474
+ "eval_accuracy": 0.606,
475
+ "eval_loss": 0.6851577162742615,
476
+ "eval_margin": 0.110994384765625,
477
+ "eval_max_reward": 1.68329541015625,
478
+ "eval_mean_reward": 1.4451339111328125,
479
+ "eval_min_reward": 1.206972412109375,
480
+ "eval_num_tokens": 480306.0,
481
+ "eval_runtime": 62.4657,
482
+ "eval_samples_per_second": 8.004,
483
+ "eval_steps_per_second": 8.004,
484
+ "step": 300
485
+ },
486
+ {
487
+ "accuracy": 0.65,
488
+ "epoch": 0.41333333333333333,
489
+ "grad_norm": 13.338096618652344,
490
+ "learning_rate": 3.97e-06,
491
+ "loss": 0.6534909725189209,
492
+ "margin": 0.14378662109375,
493
+ "max_reward": 1.83671875,
494
+ "mean_reward": 1.229180908203125,
495
+ "min_reward": 0.6009033203125,
496
+ "num_tokens": 495683.0,
497
+ "step": 310
498
+ },
499
+ {
500
+ "accuracy": 0.6,
501
+ "epoch": 0.4266666666666667,
502
+ "grad_norm": 14.492751121520996,
503
+ "learning_rate": 3.936666666666667e-06,
504
+ "loss": 0.6619410991668702,
505
+ "margin": 0.144091796875,
506
+ "max_reward": 2.00234375,
507
+ "mean_reward": 1.4490966796875,
508
+ "min_reward": 0.80947265625,
509
+ "num_tokens": 512780.0,
510
+ "step": 320
511
+ },
512
+ {
513
+ "accuracy": 0.6,
514
+ "epoch": 0.44,
515
+ "grad_norm": 17.708974838256836,
516
+ "learning_rate": 3.903333333333334e-06,
517
+ "loss": 0.6470551013946533,
518
+ "margin": 0.224090576171875,
519
+ "max_reward": 2.8828125,
520
+ "mean_reward": 2.1324859619140626,
521
+ "min_reward": 1.29166259765625,
522
+ "num_tokens": 529065.0,
523
+ "step": 330
524
+ },
525
+ {
526
+ "accuracy": 0.5,
527
+ "epoch": 0.4533333333333333,
528
+ "grad_norm": 18.278581619262695,
529
+ "learning_rate": 3.87e-06,
530
+ "loss": 0.6739425659179688,
531
+ "margin": 0.33515625,
532
+ "max_reward": 3.9375,
533
+ "mean_reward": 2.80546875,
534
+ "min_reward": 1.79609375,
535
+ "num_tokens": 546716.0,
536
+ "step": 340
537
+ },
538
+ {
539
+ "accuracy": 0.45,
540
+ "epoch": 0.4666666666666667,
541
+ "grad_norm": 38.923805236816406,
542
+ "learning_rate": 3.836666666666667e-06,
543
+ "loss": 0.7799544334411621,
544
+ "margin": -0.078515625,
545
+ "max_reward": 4.6078125,
546
+ "mean_reward": 3.6021484375,
547
+ "min_reward": 2.72265625,
548
+ "num_tokens": 563762.0,
549
+ "step": 350
550
+ },
551
+ {
552
+ "epoch": 0.4666666666666667,
553
+ "eval_accuracy": 0.572,
554
+ "eval_loss": 0.7251431345939636,
555
+ "eval_margin": 0.1519296875,
556
+ "eval_max_reward": 4.2361640625,
557
+ "eval_mean_reward": 3.866208984375,
558
+ "eval_min_reward": 3.49625390625,
559
+ "eval_num_tokens": 563762.0,
560
+ "eval_runtime": 62.2654,
561
+ "eval_samples_per_second": 8.03,
562
+ "eval_steps_per_second": 8.03,
563
+ "step": 350
564
+ },
565
+ {
566
+ "accuracy": 0.8,
567
+ "epoch": 0.48,
568
+ "grad_norm": 25.30300521850586,
569
+ "learning_rate": 3.803333333333334e-06,
570
+ "loss": 0.5838129043579101,
571
+ "margin": 0.34296875,
572
+ "max_reward": 4.70625,
573
+ "mean_reward": 3.857421875,
574
+ "min_reward": 3.0,
575
+ "num_tokens": 579498.0,
576
+ "step": 360
577
+ },
578
+ {
579
+ "accuracy": 0.5,
580
+ "epoch": 0.49333333333333335,
581
+ "grad_norm": 41.584007263183594,
582
+ "learning_rate": 3.7700000000000003e-06,
583
+ "loss": 0.8219932556152344,
584
+ "margin": 0.094140625,
585
+ "max_reward": 4.8875,
586
+ "mean_reward": 4.0658203125,
587
+ "min_reward": 3.16328125,
588
+ "num_tokens": 597098.0,
589
+ "step": 370
590
+ },
591
+ {
592
+ "accuracy": 0.65,
593
+ "epoch": 0.5066666666666667,
594
+ "grad_norm": 16.238313674926758,
595
+ "learning_rate": 3.7366666666666667e-06,
596
+ "loss": 0.6716785430908203,
597
+ "margin": 0.2427490234375,
598
+ "max_reward": 3.34375,
599
+ "mean_reward": 2.30460205078125,
600
+ "min_reward": 1.168408203125,
601
+ "num_tokens": 613640.0,
602
+ "step": 380
603
+ },
604
+ {
605
+ "accuracy": 0.6,
606
+ "epoch": 0.52,
607
+ "grad_norm": 25.986833572387695,
608
+ "learning_rate": 3.7033333333333336e-06,
609
+ "loss": 0.716916561126709,
610
+ "margin": 0.42451171875,
611
+ "max_reward": 2.96328125,
612
+ "mean_reward": 1.55986328125,
613
+ "min_reward": 0.221484375,
614
+ "num_tokens": 629441.0,
615
+ "step": 390
616
+ },
617
+ {
618
+ "accuracy": 0.75,
619
+ "epoch": 0.5333333333333333,
620
+ "grad_norm": 23.42198371887207,
621
+ "learning_rate": 3.6700000000000004e-06,
622
+ "loss": 0.6442339897155762,
623
+ "margin": 0.3801513671875,
624
+ "max_reward": 2.6359375,
625
+ "mean_reward": 1.87711181640625,
626
+ "min_reward": 1.043212890625,
627
+ "num_tokens": 645440.0,
628
+ "step": 400
629
+ },
630
+ {
631
+ "epoch": 0.5333333333333333,
632
+ "eval_accuracy": 0.618,
633
+ "eval_loss": 0.7101489305496216,
634
+ "eval_margin": 0.2907686767578125,
635
+ "eval_max_reward": 2.299792724609375,
636
+ "eval_mean_reward": 1.8211738891601563,
637
+ "eval_min_reward": 1.3425550537109374,
638
+ "eval_num_tokens": 645440.0,
639
+ "eval_runtime": 62.216,
640
+ "eval_samples_per_second": 8.037,
641
+ "eval_steps_per_second": 8.037,
642
+ "step": 400
643
+ },
644
+ {
645
+ "accuracy": 0.6,
646
+ "epoch": 0.5466666666666666,
647
+ "grad_norm": 20.917293548583984,
648
+ "learning_rate": 3.6366666666666673e-06,
649
+ "loss": 0.8052460670471191,
650
+ "margin": 0.2086181640625,
651
+ "max_reward": 2.9359375,
652
+ "mean_reward": 1.69215087890625,
653
+ "min_reward": 0.44091796875,
654
+ "num_tokens": 663799.0,
655
+ "step": 410
656
+ },
657
+ {
658
+ "accuracy": 0.7,
659
+ "epoch": 0.56,
660
+ "grad_norm": 16.333770751953125,
661
+ "learning_rate": 3.6033333333333337e-06,
662
+ "loss": 0.6448392868041992,
663
+ "margin": 0.210369873046875,
664
+ "max_reward": 0.69697265625,
665
+ "mean_reward": 0.0324127197265625,
666
+ "min_reward": -0.758642578125,
667
+ "num_tokens": 678669.0,
668
+ "step": 420
669
+ },
670
+ {
671
+ "accuracy": 0.55,
672
+ "epoch": 0.5733333333333334,
673
+ "grad_norm": 29.64971923828125,
674
+ "learning_rate": 3.57e-06,
675
+ "loss": 0.6741883277893066,
676
+ "margin": 0.1592041015625,
677
+ "max_reward": 0.2900146484375,
678
+ "mean_reward": -0.4987548828125,
679
+ "min_reward": -1.176171875,
680
+ "num_tokens": 694858.0,
681
+ "step": 430
682
+ },
683
+ {
684
+ "accuracy": 0.45,
685
+ "epoch": 0.5866666666666667,
686
+ "grad_norm": 26.162715911865234,
687
+ "learning_rate": 3.536666666666667e-06,
688
+ "loss": 0.7661936283111572,
689
+ "margin": -0.02214813232421875,
690
+ "max_reward": -0.0858306884765625,
691
+ "mean_reward": -0.7683937072753906,
692
+ "min_reward": -1.260546875,
693
+ "num_tokens": 710995.0,
694
+ "step": 440
695
+ },
696
+ {
697
+ "accuracy": 0.75,
698
+ "epoch": 0.6,
699
+ "grad_norm": 22.932537078857422,
700
+ "learning_rate": 3.5033333333333334e-06,
701
+ "loss": 0.6263980865478516,
702
+ "margin": 0.1980621337890625,
703
+ "max_reward": -0.37247314453125,
704
+ "mean_reward": -0.7766494750976562,
705
+ "min_reward": -1.234375,
706
+ "num_tokens": 726473.0,
707
+ "step": 450
708
+ },
709
+ {
710
+ "epoch": 0.6,
711
+ "eval_accuracy": 0.628,
712
+ "eval_loss": 0.6893447041511536,
713
+ "eval_margin": 0.17741348266601562,
714
+ "eval_max_reward": -0.00065301513671875,
715
+ "eval_mean_reward": -0.32534046936035155,
716
+ "eval_min_reward": -0.6500279235839844,
717
+ "eval_num_tokens": 726473.0,
718
+ "eval_runtime": 62.4229,
719
+ "eval_samples_per_second": 8.01,
720
+ "eval_steps_per_second": 8.01,
721
+ "step": 450
722
+ },
723
+ {
724
+ "accuracy": 0.45,
725
+ "epoch": 0.6133333333333333,
726
+ "grad_norm": 23.200468063354492,
727
+ "learning_rate": 3.4700000000000002e-06,
728
+ "loss": 0.7532081127166748,
729
+ "margin": -0.0011383056640625,
730
+ "max_reward": 0.3897216796875,
731
+ "mean_reward": -0.43083343505859373,
732
+ "min_reward": -1.10458984375,
733
+ "num_tokens": 743418.0,
734
+ "step": 460
735
+ },
736
+ {
737
+ "accuracy": 0.65,
738
+ "epoch": 0.6266666666666667,
739
+ "grad_norm": 23.413482666015625,
740
+ "learning_rate": 3.436666666666667e-06,
741
+ "loss": 0.6554159641265869,
742
+ "margin": 0.20365447998046876,
743
+ "max_reward": -0.0306304931640625,
744
+ "mean_reward": -0.6653846740722656,
745
+ "min_reward": -1.2474609375,
746
+ "num_tokens": 758877.0,
747
+ "step": 470
748
+ },
749
+ {
750
+ "accuracy": 0.45,
751
+ "epoch": 0.64,
752
+ "grad_norm": 23.571287155151367,
753
+ "learning_rate": 3.4033333333333335e-06,
754
+ "loss": 0.7563723564147949,
755
+ "margin": 0.1056365966796875,
756
+ "max_reward": 1.4620849609375,
757
+ "mean_reward": 0.06235504150390625,
758
+ "min_reward": -1.196258544921875,
759
+ "num_tokens": 775285.0,
760
+ "step": 480
761
+ },
762
+ {
763
+ "accuracy": 0.7,
764
+ "epoch": 0.6533333333333333,
765
+ "grad_norm": 12.891551971435547,
766
+ "learning_rate": 3.3700000000000003e-06,
767
+ "loss": 0.5850713729858399,
768
+ "margin": 0.73206787109375,
769
+ "max_reward": 2.0690673828125,
770
+ "mean_reward": 0.511407470703125,
771
+ "min_reward": -0.6796875,
772
+ "num_tokens": 790684.0,
773
+ "step": 490
774
+ },
775
+ {
776
+ "accuracy": 0.8,
777
+ "epoch": 0.6666666666666666,
778
+ "grad_norm": 64.09451293945312,
779
+ "learning_rate": 3.3366666666666668e-06,
780
+ "loss": 0.498201847076416,
781
+ "margin": 0.911553955078125,
782
+ "max_reward": 1.70078125,
783
+ "mean_reward": 0.3194183349609375,
784
+ "min_reward": -1.081884765625,
785
+ "num_tokens": 806487.0,
786
+ "step": 500
787
+ },
788
+ {
789
+ "epoch": 0.6666666666666666,
790
+ "eval_accuracy": 0.622,
791
+ "eval_loss": 0.7471536993980408,
792
+ "eval_margin": 0.40328579330444336,
793
+ "eval_max_reward": 0.7979255027770996,
794
+ "eval_mean_reward": 0.18772327995300292,
795
+ "eval_min_reward": -0.4224789428710938,
796
+ "eval_num_tokens": 806487.0,
797
+ "eval_runtime": 62.2921,
798
+ "eval_samples_per_second": 8.027,
799
+ "eval_steps_per_second": 8.027,
800
+ "step": 500
801
+ },
802
+ {
803
+ "accuracy": 0.65,
804
+ "epoch": 0.68,
805
+ "grad_norm": 122.53633117675781,
806
+ "learning_rate": 3.303333333333333e-06,
807
+ "loss": 0.7432568550109864,
808
+ "margin": 0.4142311096191406,
809
+ "max_reward": 0.76953125,
810
+ "mean_reward": -0.3576038360595703,
811
+ "min_reward": -1.263671875,
812
+ "num_tokens": 823462.0,
813
+ "step": 510
814
+ },
815
+ {
816
+ "accuracy": 0.75,
817
+ "epoch": 0.6933333333333334,
818
+ "grad_norm": 32.80317306518555,
819
+ "learning_rate": 3.2700000000000005e-06,
820
+ "loss": 0.41660318374633787,
821
+ "margin": 1.13740234375,
822
+ "max_reward": 2.3439453125,
823
+ "mean_reward": 0.789794921875,
824
+ "min_reward": -1.080078125,
825
+ "num_tokens": 838460.0,
826
+ "step": 520
827
+ },
828
+ {
829
+ "accuracy": 0.75,
830
+ "epoch": 0.7066666666666667,
831
+ "grad_norm": 17.606361389160156,
832
+ "learning_rate": 3.236666666666667e-06,
833
+ "loss": 0.6934970378875732,
834
+ "margin": 0.973486328125,
835
+ "max_reward": 4.0046875,
836
+ "mean_reward": 2.1105224609375,
837
+ "min_reward": -0.177734375,
838
+ "num_tokens": 854813.0,
839
+ "step": 530
840
+ },
841
+ {
842
+ "accuracy": 0.55,
843
+ "epoch": 0.72,
844
+ "grad_norm": 156.23760986328125,
845
+ "learning_rate": 3.2033333333333337e-06,
846
+ "loss": 1.1027880668640138,
847
+ "margin": 0.20673828125,
848
+ "max_reward": 5.353125,
849
+ "mean_reward": 3.049951171875,
850
+ "min_reward": 0.3279296875,
851
+ "num_tokens": 870996.0,
852
+ "step": 540
853
+ },
854
+ {
855
+ "accuracy": 0.45,
856
+ "epoch": 0.7333333333333333,
857
+ "grad_norm": 34.310646057128906,
858
+ "learning_rate": 3.17e-06,
859
+ "loss": 1.0296151161193847,
860
+ "margin": 0.0447265625,
861
+ "max_reward": 4.716796875,
862
+ "mean_reward": 2.67509765625,
863
+ "min_reward": 0.359375,
864
+ "num_tokens": 886715.0,
865
+ "step": 550
866
+ },
867
+ {
868
+ "epoch": 0.7333333333333333,
869
+ "eval_accuracy": 0.598,
870
+ "eval_loss": 0.8312986493110657,
871
+ "eval_margin": 0.4697259521484375,
872
+ "eval_max_reward": 3.234610900878906,
873
+ "eval_mean_reward": 2.4688244018554686,
874
+ "eval_min_reward": 1.7030379028320313,
875
+ "eval_num_tokens": 886715.0,
876
+ "eval_runtime": 62.3011,
877
+ "eval_samples_per_second": 8.026,
878
+ "eval_steps_per_second": 8.026,
879
+ "step": 550
880
+ },
881
+ {
882
+ "accuracy": 0.65,
883
+ "epoch": 0.7466666666666667,
884
+ "grad_norm": 2.1129918098449707,
885
+ "learning_rate": 3.1366666666666666e-06,
886
+ "loss": 0.6547252178192139,
887
+ "margin": 0.72010498046875,
888
+ "max_reward": 3.6328125,
889
+ "mean_reward": 1.748907470703125,
890
+ "min_reward": -0.1922607421875,
891
+ "num_tokens": 901702.0,
892
+ "step": 560
893
+ },
894
+ {
895
+ "accuracy": 0.55,
896
+ "epoch": 0.76,
897
+ "grad_norm": 19.6908016204834,
898
+ "learning_rate": 3.103333333333334e-06,
899
+ "loss": 0.6725973129272461,
900
+ "margin": 0.536669921875,
901
+ "max_reward": 0.89931640625,
902
+ "mean_reward": -0.4901123046875,
903
+ "min_reward": -1.93125,
904
+ "num_tokens": 919528.0,
905
+ "step": 570
906
+ },
907
+ {
908
+ "accuracy": 0.5,
909
+ "epoch": 0.7733333333333333,
910
+ "grad_norm": 56.94779586791992,
911
+ "learning_rate": 3.0700000000000003e-06,
912
+ "loss": 0.6944450378417969,
913
+ "margin": 0.448974609375,
914
+ "max_reward": 0.27529296875,
915
+ "mean_reward": -1.3425048828125,
916
+ "min_reward": -2.58046875,
917
+ "num_tokens": 936620.0,
918
+ "step": 580
919
+ },
920
+ {
921
+ "accuracy": 0.85,
922
+ "epoch": 0.7866666666666666,
923
+ "grad_norm": 95.96953582763672,
924
+ "learning_rate": 3.0366666666666667e-06,
925
+ "loss": 0.3615936040878296,
926
+ "margin": 1.50234375,
927
+ "max_reward": 1.0953125,
928
+ "mean_reward": -1.33935546875,
929
+ "min_reward": -3.105078125,
930
+ "num_tokens": 953455.0,
931
+ "step": 590
932
+ },
933
+ {
934
+ "accuracy": 0.7,
935
+ "epoch": 0.8,
936
+ "grad_norm": 88.94745635986328,
937
+ "learning_rate": 3.0033333333333335e-06,
938
+ "loss": 0.9391720771789551,
939
+ "margin": 1.326025390625,
940
+ "max_reward": 5.046875,
941
+ "mean_reward": 2.1633544921875,
942
+ "min_reward": -1.42548828125,
943
+ "num_tokens": 968502.0,
944
+ "step": 600
945
+ },
946
+ {
947
+ "epoch": 0.8,
948
+ "eval_accuracy": 0.598,
949
+ "eval_loss": 1.15407395362854,
950
+ "eval_margin": 0.8387359619140625,
951
+ "eval_max_reward": 3.26913037109375,
952
+ "eval_mean_reward": 1.9263014526367188,
953
+ "eval_min_reward": 0.5834725341796875,
954
+ "eval_num_tokens": 968502.0,
955
+ "eval_runtime": 62.2605,
956
+ "eval_samples_per_second": 8.031,
957
+ "eval_steps_per_second": 8.031,
958
+ "step": 600
959
+ },
960
+ {
961
+ "accuracy": 0.5,
962
+ "epoch": 0.8133333333333334,
963
+ "grad_norm": 32.30426025390625,
964
+ "learning_rate": 2.97e-06,
965
+ "loss": 1.0049471855163574,
966
+ "margin": 0.078076171875,
967
+ "max_reward": 4.67255859375,
968
+ "mean_reward": 1.0933349609375,
969
+ "min_reward": -2.70625,
970
+ "num_tokens": 985736.0,
971
+ "step": 610
972
+ },
973
+ {
974
+ "accuracy": 0.7,
975
+ "epoch": 0.8266666666666667,
976
+ "grad_norm": 51.47792434692383,
977
+ "learning_rate": 2.936666666666667e-06,
978
+ "loss": 0.7427908420562744,
979
+ "margin": 0.39453125,
980
+ "max_reward": 5.634375,
981
+ "mean_reward": 4.305078125,
982
+ "min_reward": 2.32734375,
983
+ "num_tokens": 1002008.0,
984
+ "step": 620
985
+ },
986
+ {
987
+ "accuracy": 0.5,
988
+ "epoch": 0.84,
989
+ "grad_norm": 23.034889221191406,
990
+ "learning_rate": 2.9033333333333336e-06,
991
+ "loss": 0.8153639793395996,
992
+ "margin": -0.12333984375,
993
+ "max_reward": 5.821875,
994
+ "mean_reward": 4.691064453125,
995
+ "min_reward": 3.4435546875,
996
+ "num_tokens": 1019028.0,
997
+ "step": 630
998
+ },
999
+ {
1000
+ "accuracy": 0.35,
1001
+ "epoch": 0.8533333333333334,
1002
+ "grad_norm": 161.87069702148438,
1003
+ "learning_rate": 2.87e-06,
1004
+ "loss": 0.8048789978027344,
1005
+ "margin": -0.009716796875,
1006
+ "max_reward": 5.628125,
1007
+ "mean_reward": 4.9048583984375,
1008
+ "min_reward": 3.97490234375,
1009
+ "num_tokens": 1033366.0,
1010
+ "step": 640
1011
+ },
1012
+ {
1013
+ "accuracy": 0.55,
1014
+ "epoch": 0.8666666666666667,
1015
+ "grad_norm": 15.923206329345703,
1016
+ "learning_rate": 2.836666666666667e-06,
1017
+ "loss": 0.7137910842895507,
1018
+ "margin": 0.00390625,
1019
+ "max_reward": 5.36875,
1020
+ "mean_reward": 4.934765625,
1021
+ "min_reward": 4.5046875,
1022
+ "num_tokens": 1050252.0,
1023
+ "step": 650
1024
+ },
1025
+ {
1026
+ "epoch": 0.8666666666666667,
1027
+ "eval_accuracy": 0.526,
1028
+ "eval_loss": 0.7010688185691833,
1029
+ "eval_margin": 0.0384140625,
1030
+ "eval_max_reward": 5.014875,
1031
+ "eval_mean_reward": 4.85937109375,
1032
+ "eval_min_reward": 4.7038671875,
1033
+ "eval_num_tokens": 1050252.0,
1034
+ "eval_runtime": 62.339,
1035
+ "eval_samples_per_second": 8.021,
1036
+ "eval_steps_per_second": 8.021,
1037
+ "step": 650
1038
+ },
1039
+ {
1040
+ "accuracy": 0.6,
1041
+ "epoch": 0.88,
1042
+ "grad_norm": 15.554156303405762,
1043
+ "learning_rate": 2.8033333333333333e-06,
1044
+ "loss": 0.65819411277771,
1045
+ "margin": 0.16484375,
1046
+ "max_reward": 5.25,
1047
+ "mean_reward": 4.766015625,
1048
+ "min_reward": 4.1578125,
1049
+ "num_tokens": 1068961.0,
1050
+ "step": 660
1051
+ },
1052
+ {
1053
+ "accuracy": 0.55,
1054
+ "epoch": 0.8933333333333333,
1055
+ "grad_norm": 14.689868927001953,
1056
+ "learning_rate": 2.7700000000000006e-06,
1057
+ "loss": 0.68900465965271,
1058
+ "margin": 0.12265625,
1059
+ "max_reward": 4.746875,
1060
+ "mean_reward": 4.285546875,
1061
+ "min_reward": 3.6515625,
1062
+ "num_tokens": 1084178.0,
1063
+ "step": 670
1064
+ },
1065
+ {
1066
+ "accuracy": 0.65,
1067
+ "epoch": 0.9066666666666666,
1068
+ "grad_norm": 30.876516342163086,
1069
+ "learning_rate": 2.736666666666667e-06,
1070
+ "loss": 0.6992377281188965,
1071
+ "margin": 0.0375,
1072
+ "max_reward": 4.3859375,
1073
+ "mean_reward": 4.02265625,
1074
+ "min_reward": 3.6703125,
1075
+ "num_tokens": 1100910.0,
1076
+ "step": 680
1077
+ },
1078
+ {
1079
+ "accuracy": 0.65,
1080
+ "epoch": 0.92,
1081
+ "grad_norm": 24.66133689880371,
1082
+ "learning_rate": 2.7033333333333334e-06,
1083
+ "loss": 0.6626120567321777,
1084
+ "margin": 0.20625,
1085
+ "max_reward": 4.084375,
1086
+ "mean_reward": 3.6171875,
1087
+ "min_reward": 2.8875,
1088
+ "num_tokens": 1116440.0,
1089
+ "step": 690
1090
+ },
1091
+ {
1092
+ "accuracy": 0.45,
1093
+ "epoch": 0.9333333333333333,
1094
+ "grad_norm": 15.120383262634277,
1095
+ "learning_rate": 2.6700000000000003e-06,
1096
+ "loss": 0.696980619430542,
1097
+ "margin": 0.13203125,
1098
+ "max_reward": 4.3109375,
1099
+ "mean_reward": 3.649609375,
1100
+ "min_reward": 3.0734375,
1101
+ "num_tokens": 1132429.0,
1102
+ "step": 700
1103
+ },
1104
+ {
1105
+ "epoch": 0.9333333333333333,
1106
+ "eval_accuracy": 0.582,
1107
+ "eval_loss": 0.7070507407188416,
1108
+ "eval_margin": 0.0835546875,
1109
+ "eval_max_reward": 3.954890625,
1110
+ "eval_mean_reward": 3.70844921875,
1111
+ "eval_min_reward": 3.4620078125,
1112
+ "eval_num_tokens": 1132429.0,
1113
+ "eval_runtime": 62.4281,
1114
+ "eval_samples_per_second": 8.009,
1115
+ "eval_steps_per_second": 8.009,
1116
+ "step": 700
1117
+ },
1118
+ {
1119
+ "accuracy": 0.55,
1120
+ "epoch": 0.9466666666666667,
1121
+ "grad_norm": 14.82724380493164,
1122
+ "learning_rate": 2.6366666666666667e-06,
1123
+ "loss": 0.6956966400146485,
1124
+ "margin": 0.065625,
1125
+ "max_reward": 4.340625,
1126
+ "mean_reward": 3.9,
1127
+ "min_reward": 3.3796875,
1128
+ "num_tokens": 1148620.0,
1129
+ "step": 710
1130
+ },
1131
+ {
1132
+ "accuracy": 0.35,
1133
+ "epoch": 0.96,
1134
+ "grad_norm": 28.879314422607422,
1135
+ "learning_rate": 2.603333333333334e-06,
1136
+ "loss": 0.7410328388214111,
1137
+ "margin": -0.046875,
1138
+ "max_reward": 4.2875,
1139
+ "mean_reward": 3.8375,
1140
+ "min_reward": 3.3515625,
1141
+ "num_tokens": 1164601.0,
1142
+ "step": 720
1143
+ },
1144
+ {
1145
+ "accuracy": 0.65,
1146
+ "epoch": 0.9733333333333334,
1147
+ "grad_norm": 17.8446044921875,
1148
+ "learning_rate": 2.5700000000000004e-06,
1149
+ "loss": 0.6587936878204346,
1150
+ "margin": 0.2615234375,
1151
+ "max_reward": 4.5078125,
1152
+ "mean_reward": 3.79892578125,
1153
+ "min_reward": 3.089453125,
1154
+ "num_tokens": 1179554.0,
1155
+ "step": 730
1156
+ },
1157
+ {
1158
+ "accuracy": 0.55,
1159
+ "epoch": 0.9866666666666667,
1160
+ "grad_norm": 16.74919891357422,
1161
+ "learning_rate": 2.536666666666667e-06,
1162
+ "loss": 0.9084688186645508,
1163
+ "margin": -0.1103515625,
1164
+ "max_reward": 3.65,
1165
+ "mean_reward": 2.99423828125,
1166
+ "min_reward": 1.926953125,
1167
+ "num_tokens": 1195986.0,
1168
+ "step": 740
1169
+ },
1170
+ {
1171
+ "accuracy": 0.6,
1172
+ "epoch": 1.0,
1173
+ "grad_norm": 20.36426544189453,
1174
+ "learning_rate": 2.5033333333333336e-06,
1175
+ "loss": 0.6503221035003662,
1176
+ "margin": 0.196875,
1177
+ "max_reward": 3.8296875,
1178
+ "mean_reward": 3.22890625,
1179
+ "min_reward": 2.6796875,
1180
+ "num_tokens": 1212807.0,
1181
+ "step": 750
1182
+ },
1183
+ {
1184
+ "epoch": 1.0,
1185
+ "eval_accuracy": 0.574,
1186
+ "eval_loss": 0.724712073802948,
1187
+ "eval_margin": 0.1419578857421875,
1188
+ "eval_max_reward": 3.170109375,
1189
+ "eval_mean_reward": 2.827115661621094,
1190
+ "eval_min_reward": 2.4841219482421875,
1191
+ "eval_num_tokens": 1212807.0,
1192
+ "eval_runtime": 62.4609,
1193
+ "eval_samples_per_second": 8.005,
1194
+ "eval_steps_per_second": 8.005,
1195
+ "step": 750
1196
+ },
1197
+ {
1198
+ "accuracy": 0.8,
1199
+ "epoch": 1.0133333333333334,
1200
+ "grad_norm": 25.44012451171875,
1201
+ "learning_rate": 2.47e-06,
1202
+ "loss": 0.5082518577575683,
1203
+ "margin": 0.634228515625,
1204
+ "max_reward": 3.65,
1205
+ "mean_reward": 2.5745849609375,
1206
+ "min_reward": 1.33837890625,
1207
+ "num_tokens": 1230748.0,
1208
+ "step": 760
1209
+ },
1210
+ {
1211
+ "accuracy": 0.6,
1212
+ "epoch": 1.0266666666666666,
1213
+ "grad_norm": 18.013717651367188,
1214
+ "learning_rate": 2.436666666666667e-06,
1215
+ "loss": 0.5603197574615478,
1216
+ "margin": 0.5654541015625,
1217
+ "max_reward": 3.029833984375,
1218
+ "mean_reward": 1.75579833984375,
1219
+ "min_reward": 0.37646484375,
1220
+ "num_tokens": 1245628.0,
1221
+ "step": 770
1222
+ },
1223
+ {
1224
+ "accuracy": 0.55,
1225
+ "epoch": 1.04,
1226
+ "grad_norm": 99.6661148071289,
1227
+ "learning_rate": 2.4033333333333338e-06,
1228
+ "loss": 0.8869245529174805,
1229
+ "margin": 0.205218505859375,
1230
+ "max_reward": 2.080810546875,
1231
+ "mean_reward": -0.0414886474609375,
1232
+ "min_reward": -2.2107421875,
1233
+ "num_tokens": 1262533.0,
1234
+ "step": 780
1235
+ },
1236
+ {
1237
+ "accuracy": 0.7,
1238
+ "epoch": 1.0533333333333332,
1239
+ "grad_norm": 53.62653732299805,
1240
+ "learning_rate": 2.37e-06,
1241
+ "loss": 0.5512317657470703,
1242
+ "margin": 1.0128662109375,
1243
+ "max_reward": 1.4505859375,
1244
+ "mean_reward": -0.33284912109375,
1245
+ "min_reward": -2.0107421875,
1246
+ "num_tokens": 1277096.0,
1247
+ "step": 790
1248
+ },
1249
+ {
1250
+ "accuracy": 0.7,
1251
+ "epoch": 1.0666666666666667,
1252
+ "grad_norm": 95.58021545410156,
1253
+ "learning_rate": 2.3366666666666666e-06,
1254
+ "loss": 0.6172554016113281,
1255
+ "margin": 1.2177459716796875,
1256
+ "max_reward": 1.304345703125,
1257
+ "mean_reward": -1.1034286499023438,
1258
+ "min_reward": -3.26953125,
1259
+ "num_tokens": 1293073.0,
1260
+ "step": 800
1261
+ },
1262
+ {
1263
+ "epoch": 1.0666666666666667,
1264
+ "eval_accuracy": 0.596,
1265
+ "eval_loss": 0.8903509378433228,
1266
+ "eval_margin": 0.564534423828125,
1267
+ "eval_max_reward": -0.50098583984375,
1268
+ "eval_mean_reward": -1.4187645263671875,
1269
+ "eval_min_reward": -2.336543212890625,
1270
+ "eval_num_tokens": 1293073.0,
1271
+ "eval_runtime": 62.3781,
1272
+ "eval_samples_per_second": 8.016,
1273
+ "eval_steps_per_second": 8.016,
1274
+ "step": 800
1275
+ },
1276
+ {
1277
+ "accuracy": 0.65,
1278
+ "epoch": 1.08,
1279
+ "grad_norm": 34.59308624267578,
1280
+ "learning_rate": 2.3033333333333334e-06,
1281
+ "loss": 0.7249170303344726,
1282
+ "margin": 0.968017578125,
1283
+ "max_reward": 0.56064453125,
1284
+ "mean_reward": -1.4824951171875,
1285
+ "min_reward": -3.8390625,
1286
+ "num_tokens": 1308687.0,
1287
+ "step": 810
1288
+ },
1289
+ {
1290
+ "accuracy": 0.8,
1291
+ "epoch": 1.0933333333333333,
1292
+ "grad_norm": 30.92959976196289,
1293
+ "learning_rate": 2.2700000000000003e-06,
1294
+ "loss": 0.39841785430908205,
1295
+ "margin": 1.375030517578125,
1296
+ "max_reward": 1.658984375,
1297
+ "mean_reward": -0.6310943603515625,
1298
+ "min_reward": -3.046484375,
1299
+ "num_tokens": 1324618.0,
1300
+ "step": 820
1301
+ },
1302
+ {
1303
+ "accuracy": 0.7,
1304
+ "epoch": 1.1066666666666667,
1305
+ "grad_norm": 208.02574157714844,
1306
+ "learning_rate": 2.236666666666667e-06,
1307
+ "loss": 0.8557153701782226,
1308
+ "margin": 1.5063720703125,
1309
+ "max_reward": 1.581640625,
1310
+ "mean_reward": -0.71844482421875,
1311
+ "min_reward": -3.421484375,
1312
+ "num_tokens": 1339988.0,
1313
+ "step": 830
1314
+ },
1315
+ {
1316
+ "accuracy": 0.75,
1317
+ "epoch": 1.12,
1318
+ "grad_norm": 279.54534912109375,
1319
+ "learning_rate": 2.2033333333333336e-06,
1320
+ "loss": 0.7527968406677246,
1321
+ "margin": 2.487060546875,
1322
+ "max_reward": 1.5951171875,
1323
+ "mean_reward": -2.3231201171875,
1324
+ "min_reward": -5.828125,
1325
+ "num_tokens": 1356112.0,
1326
+ "step": 840
1327
+ },
1328
+ {
1329
+ "accuracy": 0.7,
1330
+ "epoch": 1.1333333333333333,
1331
+ "grad_norm": 35.76692581176758,
1332
+ "learning_rate": 2.17e-06,
1333
+ "loss": 0.8942340850830078,
1334
+ "margin": 1.6185546875,
1335
+ "max_reward": 0.787890625,
1336
+ "mean_reward": -1.966064453125,
1337
+ "min_reward": -5.40625,
1338
+ "num_tokens": 1372383.0,
1339
+ "step": 850
1340
+ },
1341
+ {
1342
+ "epoch": 1.1333333333333333,
1343
+ "eval_accuracy": 0.596,
1344
+ "eval_loss": 1.0435729026794434,
1345
+ "eval_margin": 0.8098057250976562,
1346
+ "eval_max_reward": 0.7392849731445312,
1347
+ "eval_mean_reward": -0.5019579772949219,
1348
+ "eval_min_reward": -1.743200927734375,
1349
+ "eval_num_tokens": 1372383.0,
1350
+ "eval_runtime": 62.4519,
1351
+ "eval_samples_per_second": 8.006,
1352
+ "eval_steps_per_second": 8.006,
1353
+ "step": 850
1354
+ },
1355
+ {
1356
+ "accuracy": 0.8,
1357
+ "epoch": 1.1466666666666667,
1358
+ "grad_norm": 18.71978759765625,
1359
+ "learning_rate": 2.136666666666667e-06,
1360
+ "loss": 0.6701197147369384,
1361
+ "margin": 1.843017578125,
1362
+ "max_reward": 3.765625,
1363
+ "mean_reward": 0.3057373046875,
1364
+ "min_reward": -3.39443359375,
1365
+ "num_tokens": 1388575.0,
1366
+ "step": 860
1367
+ },
1368
+ {
1369
+ "accuracy": 0.7,
1370
+ "epoch": 1.16,
1371
+ "grad_norm": 150.29055786132812,
1372
+ "learning_rate": 2.1033333333333337e-06,
1373
+ "loss": 0.6811777114868164,
1374
+ "margin": 1.3685546875,
1375
+ "max_reward": 2.4873046875,
1376
+ "mean_reward": 0.19677734375,
1377
+ "min_reward": -2.523046875,
1378
+ "num_tokens": 1403606.0,
1379
+ "step": 870
1380
+ },
1381
+ {
1382
+ "accuracy": 0.75,
1383
+ "epoch": 1.1733333333333333,
1384
+ "grad_norm": 187.19081115722656,
1385
+ "learning_rate": 2.07e-06,
1386
+ "loss": 0.8808813095092773,
1387
+ "margin": 1.917291259765625,
1388
+ "max_reward": 4.8203125,
1389
+ "mean_reward": 2.1001434326171875,
1390
+ "min_reward": -1.4076171875,
1391
+ "num_tokens": 1418488.0,
1392
+ "step": 880
1393
+ },
1394
+ {
1395
+ "accuracy": 0.8,
1396
+ "epoch": 1.1866666666666668,
1397
+ "grad_norm": 44.864192962646484,
1398
+ "learning_rate": 2.036666666666667e-06,
1399
+ "loss": 0.558538818359375,
1400
+ "margin": 2.50665283203125,
1401
+ "max_reward": 4.6734375,
1402
+ "mean_reward": 1.906243896484375,
1403
+ "min_reward": -2.1400634765625,
1404
+ "num_tokens": 1433529.0,
1405
+ "step": 890
1406
+ },
1407
+ {
1408
+ "accuracy": 0.8,
1409
+ "epoch": 1.2,
1410
+ "grad_norm": 46.679744720458984,
1411
+ "learning_rate": 2.0033333333333334e-06,
1412
+ "loss": 0.7218273639678955,
1413
+ "margin": 2.88583984375,
1414
+ "max_reward": 3.57421875,
1415
+ "mean_reward": -0.101611328125,
1416
+ "min_reward": -4.85234375,
1417
+ "num_tokens": 1450110.0,
1418
+ "step": 900
1419
+ },
1420
+ {
1421
+ "epoch": 1.2,
1422
+ "eval_accuracy": 0.608,
1423
+ "eval_loss": 1.363513708114624,
1424
+ "eval_margin": 1.084634536743164,
1425
+ "eval_max_reward": 3.07980419921875,
1426
+ "eval_mean_reward": 1.359497184753418,
1427
+ "eval_min_reward": -0.36080982971191405,
1428
+ "eval_num_tokens": 1450110.0,
1429
+ "eval_runtime": 62.4495,
1430
+ "eval_samples_per_second": 8.006,
1431
+ "eval_steps_per_second": 8.006,
1432
+ "step": 900
1433
+ },
1434
+ {
1435
+ "accuracy": 0.6,
1436
+ "epoch": 1.2133333333333334,
1437
+ "grad_norm": 190.0553741455078,
1438
+ "learning_rate": 1.97e-06,
1439
+ "loss": 0.7698217391967773,
1440
+ "margin": 1.106640625,
1441
+ "max_reward": 5.125,
1442
+ "mean_reward": 1.506640625,
1443
+ "min_reward": -2.765625,
1444
+ "num_tokens": 1466027.0,
1445
+ "step": 910
1446
+ },
1447
+ {
1448
+ "accuracy": 0.6,
1449
+ "epoch": 1.2266666666666666,
1450
+ "grad_norm": 234.94076538085938,
1451
+ "learning_rate": 1.9366666666666666e-06,
1452
+ "loss": 1.6000349044799804,
1453
+ "margin": 1.884661865234375,
1454
+ "max_reward": 1.996875,
1455
+ "mean_reward": -2.3544403076171876,
1456
+ "min_reward": -8.725,
1457
+ "num_tokens": 1481748.0,
1458
+ "step": 920
1459
+ },
1460
+ {
1461
+ "accuracy": 0.55,
1462
+ "epoch": 1.24,
1463
+ "grad_norm": 65.35015869140625,
1464
+ "learning_rate": 1.9033333333333335e-06,
1465
+ "loss": 1.3991120338439942,
1466
+ "margin": 1.012890625,
1467
+ "max_reward": -1.74296875,
1468
+ "mean_reward": -5.5919921875,
1469
+ "min_reward": -9.453125,
1470
+ "num_tokens": 1498020.0,
1471
+ "step": 930
1472
+ },
1473
+ {
1474
+ "accuracy": 0.75,
1475
+ "epoch": 1.2533333333333334,
1476
+ "grad_norm": 57.545166015625,
1477
+ "learning_rate": 1.87e-06,
1478
+ "loss": 0.49834294319152833,
1479
+ "margin": 2.0171875,
1480
+ "max_reward": -3.440625,
1481
+ "mean_reward": -6.62265625,
1482
+ "min_reward": -9.015625,
1483
+ "num_tokens": 1514583.0,
1484
+ "step": 940
1485
+ },
1486
+ {
1487
+ "accuracy": 0.85,
1488
+ "epoch": 1.2666666666666666,
1489
+ "grad_norm": 144.95867919921875,
1490
+ "learning_rate": 1.836666666666667e-06,
1491
+ "loss": 0.5361029624938964,
1492
+ "margin": 3.177197265625,
1493
+ "max_reward": 0.4771484375,
1494
+ "mean_reward": -3.9238525390625,
1495
+ "min_reward": -8.1375,
1496
+ "num_tokens": 1530955.0,
1497
+ "step": 950
1498
+ },
1499
+ {
1500
+ "epoch": 1.2666666666666666,
1501
+ "eval_accuracy": 0.596,
1502
+ "eval_loss": 1.3600239753723145,
1503
+ "eval_margin": 1.1310927734375,
1504
+ "eval_max_reward": -1.742294677734375,
1505
+ "eval_mean_reward": -3.521328125,
1506
+ "eval_min_reward": -5.300361572265625,
1507
+ "eval_num_tokens": 1530955.0,
1508
+ "eval_runtime": 62.6012,
1509
+ "eval_samples_per_second": 7.987,
1510
+ "eval_steps_per_second": 7.987,
1511
+ "step": 950
1512
+ },
1513
+ {
1514
+ "accuracy": 0.5,
1515
+ "epoch": 1.28,
1516
+ "grad_norm": 319.9293518066406,
1517
+ "learning_rate": 1.8033333333333336e-06,
1518
+ "loss": 1.7453046798706056,
1519
+ "margin": 0.3447265625,
1520
+ "max_reward": 1.01591796875,
1521
+ "mean_reward": -3.259765625,
1522
+ "min_reward": -7.2578125,
1523
+ "num_tokens": 1548332.0,
1524
+ "step": 960
1525
+ },
1526
+ {
1527
+ "accuracy": 0.65,
1528
+ "epoch": 1.2933333333333334,
1529
+ "grad_norm": 33.143341064453125,
1530
+ "learning_rate": 1.77e-06,
1531
+ "loss": 0.7040543556213379,
1532
+ "margin": 2.181396484375,
1533
+ "max_reward": -0.30205078125,
1534
+ "mean_reward": -3.7452392578125,
1535
+ "min_reward": -6.275,
1536
+ "num_tokens": 1564625.0,
1537
+ "step": 970
1538
+ },
1539
+ {
1540
+ "accuracy": 0.75,
1541
+ "epoch": 1.3066666666666666,
1542
+ "grad_norm": 31.115459442138672,
1543
+ "learning_rate": 1.7366666666666668e-06,
1544
+ "loss": 0.4173430919647217,
1545
+ "margin": 2.164306640625,
1546
+ "max_reward": 2.28095703125,
1547
+ "mean_reward": -0.9407470703125,
1548
+ "min_reward": -4.04375,
1549
+ "num_tokens": 1580000.0,
1550
+ "step": 980
1551
+ },
1552
+ {
1553
+ "accuracy": 0.8,
1554
+ "epoch": 1.32,
1555
+ "grad_norm": 168.20628356933594,
1556
+ "learning_rate": 1.7033333333333335e-06,
1557
+ "loss": 0.5366679191589355,
1558
+ "margin": 3.0508544921875,
1559
+ "max_reward": 2.8671875,
1560
+ "mean_reward": -1.78248291015625,
1561
+ "min_reward": -5.2830078125,
1562
+ "num_tokens": 1597302.0,
1563
+ "step": 990
1564
+ },
1565
+ {
1566
+ "accuracy": 0.45,
1567
+ "epoch": 1.3333333333333333,
1568
+ "grad_norm": 18.123050689697266,
1569
+ "learning_rate": 1.6700000000000003e-06,
1570
+ "loss": 1.1453736305236817,
1571
+ "margin": 0.8822265625,
1572
+ "max_reward": 4.8984375,
1573
+ "mean_reward": 2.26845703125,
1574
+ "min_reward": -0.678515625,
1575
+ "num_tokens": 1613437.0,
1576
+ "step": 1000
1577
+ },
1578
+ {
1579
+ "epoch": 1.3333333333333333,
1580
+ "eval_accuracy": 0.586,
1581
+ "eval_loss": 1.1093744039535522,
1582
+ "eval_margin": 0.682355224609375,
1583
+ "eval_max_reward": 3.630033935546875,
1584
+ "eval_mean_reward": 2.4088892822265624,
1585
+ "eval_min_reward": 1.18774462890625,
1586
+ "eval_num_tokens": 1613437.0,
1587
+ "eval_runtime": 62.4977,
1588
+ "eval_samples_per_second": 8.0,
1589
+ "eval_steps_per_second": 8.0,
1590
+ "step": 1000
1591
+ },
1592
+ {
1593
+ "accuracy": 0.75,
1594
+ "epoch": 1.3466666666666667,
1595
+ "grad_norm": 35.67937088012695,
1596
+ "learning_rate": 1.6366666666666667e-06,
1597
+ "loss": 0.675491189956665,
1598
+ "margin": 0.5638671875,
1599
+ "max_reward": 4.290625,
1600
+ "mean_reward": 2.44892578125,
1601
+ "min_reward": 0.67578125,
1602
+ "num_tokens": 1629455.0,
1603
+ "step": 1010
1604
+ },
1605
+ {
1606
+ "accuracy": 0.7,
1607
+ "epoch": 1.3599999999999999,
1608
+ "grad_norm": 40.7118034362793,
1609
+ "learning_rate": 1.6033333333333334e-06,
1610
+ "loss": 0.42725315093994143,
1611
+ "margin": 1.8513671875,
1612
+ "max_reward": 4.996875,
1613
+ "mean_reward": 3.25732421875,
1614
+ "min_reward": 0.50625,
1615
+ "num_tokens": 1644785.0,
1616
+ "step": 1020
1617
+ },
1618
+ {
1619
+ "accuracy": 0.8,
1620
+ "epoch": 1.3733333333333333,
1621
+ "grad_norm": 136.8484344482422,
1622
+ "learning_rate": 1.5700000000000002e-06,
1623
+ "loss": 0.6001864433288574,
1624
+ "margin": 3.10028076171875,
1625
+ "max_reward": 4.1550537109375,
1626
+ "mean_reward": 0.757366943359375,
1627
+ "min_reward": -3.25009765625,
1628
+ "num_tokens": 1661544.0,
1629
+ "step": 1030
1630
+ },
1631
+ {
1632
+ "accuracy": 0.75,
1633
+ "epoch": 1.3866666666666667,
1634
+ "grad_norm": 0.3609694242477417,
1635
+ "learning_rate": 1.5366666666666668e-06,
1636
+ "loss": 0.9377962112426758,
1637
+ "margin": 1.72666015625,
1638
+ "max_reward": 3.7875,
1639
+ "mean_reward": 1.237353515625,
1640
+ "min_reward": -2.774609375,
1641
+ "num_tokens": 1677000.0,
1642
+ "step": 1040
1643
+ },
1644
+ {
1645
+ "accuracy": 0.7,
1646
+ "epoch": 1.4,
1647
+ "grad_norm": 16.80458641052246,
1648
+ "learning_rate": 1.5033333333333337e-06,
1649
+ "loss": 0.7020199298858643,
1650
+ "margin": 1.91923828125,
1651
+ "max_reward": 3.17734375,
1652
+ "mean_reward": 0.354052734375,
1653
+ "min_reward": -3.8123046875,
1654
+ "num_tokens": 1691904.0,
1655
+ "step": 1050
1656
+ },
1657
+ {
1658
+ "epoch": 1.4,
1659
+ "eval_accuracy": 0.606,
1660
+ "eval_loss": 1.3567456007003784,
1661
+ "eval_margin": 1.0848167724609374,
1662
+ "eval_max_reward": 1.356125732421875,
1663
+ "eval_mean_reward": -0.37665606689453124,
1664
+ "eval_min_reward": -2.1094378662109374,
1665
+ "eval_num_tokens": 1691904.0,
1666
+ "eval_runtime": 62.3936,
1667
+ "eval_samples_per_second": 8.014,
1668
+ "eval_steps_per_second": 8.014,
1669
+ "step": 1050
1670
+ },
1671
+ {
1672
+ "accuracy": 0.55,
1673
+ "epoch": 1.4133333333333333,
1674
+ "grad_norm": 117.69019317626953,
1675
+ "learning_rate": 1.4700000000000001e-06,
1676
+ "loss": 1.212506103515625,
1677
+ "margin": 1.3871475219726563,
1678
+ "max_reward": 2.3145721435546873,
1679
+ "mean_reward": -3.046747589111328,
1680
+ "min_reward": -6.746875,
1681
+ "num_tokens": 1708299.0,
1682
+ "step": 1060
1683
+ },
1684
+ {
1685
+ "accuracy": 0.6,
1686
+ "epoch": 1.4266666666666667,
1687
+ "grad_norm": 100.34335327148438,
1688
+ "learning_rate": 1.4366666666666667e-06,
1689
+ "loss": 0.9856803894042969,
1690
+ "margin": 1.3234375,
1691
+ "max_reward": 1.334375,
1692
+ "mean_reward": -2.6033203125,
1693
+ "min_reward": -6.49765625,
1694
+ "num_tokens": 1725007.0,
1695
+ "step": 1070
1696
+ },
1697
+ {
1698
+ "accuracy": 0.7,
1699
+ "epoch": 1.44,
1700
+ "grad_norm": 185.3729248046875,
1701
+ "learning_rate": 1.4033333333333336e-06,
1702
+ "loss": 1.1272445678710938,
1703
+ "margin": 2.0210784912109374,
1704
+ "max_reward": 3.303125,
1705
+ "mean_reward": -0.05469512939453125,
1706
+ "min_reward": -3.634649658203125,
1707
+ "num_tokens": 1740533.0,
1708
+ "step": 1080
1709
+ },
1710
+ {
1711
+ "accuracy": 0.6,
1712
+ "epoch": 1.4533333333333334,
1713
+ "grad_norm": 161.89877319335938,
1714
+ "learning_rate": 1.3700000000000002e-06,
1715
+ "loss": 1.4939114570617675,
1716
+ "margin": 0.5677734375,
1717
+ "max_reward": 1.749609375,
1718
+ "mean_reward": -2.25419921875,
1719
+ "min_reward": -6.180859375,
1720
+ "num_tokens": 1757099.0,
1721
+ "step": 1090
1722
+ },
1723
+ {
1724
+ "accuracy": 0.75,
1725
+ "epoch": 1.4666666666666668,
1726
+ "grad_norm": 119.40484619140625,
1727
+ "learning_rate": 1.3366666666666666e-06,
1728
+ "loss": 0.8826108932495117,
1729
+ "margin": 1.85439453125,
1730
+ "max_reward": 2.553125,
1731
+ "mean_reward": -0.852197265625,
1732
+ "min_reward": -4.6171875,
1733
+ "num_tokens": 1775146.0,
1734
+ "step": 1100
1735
+ },
1736
+ {
1737
+ "epoch": 1.4666666666666668,
1738
+ "eval_accuracy": 0.604,
1739
+ "eval_loss": 0.9978691339492798,
1740
+ "eval_margin": 0.7300045776367188,
1741
+ "eval_max_reward": 0.6507423095703125,
1742
+ "eval_mean_reward": -0.4976886901855469,
1743
+ "eval_min_reward": -1.6461196899414063,
1744
+ "eval_num_tokens": 1775146.0,
1745
+ "eval_runtime": 62.4188,
1746
+ "eval_samples_per_second": 8.01,
1747
+ "eval_steps_per_second": 8.01,
1748
+ "step": 1100
1749
+ },
1750
+ {
1751
+ "accuracy": 0.65,
1752
+ "epoch": 1.48,
1753
+ "grad_norm": 109.5771713256836,
1754
+ "learning_rate": 1.3033333333333335e-06,
1755
+ "loss": 0.9664765357971191,
1756
+ "margin": 0.737762451171875,
1757
+ "max_reward": 1.653662109375,
1758
+ "mean_reward": -0.5663665771484375,
1759
+ "min_reward": -3.6734375,
1760
+ "num_tokens": 1793063.0,
1761
+ "step": 1110
1762
+ },
1763
+ {
1764
+ "accuracy": 0.7,
1765
+ "epoch": 1.4933333333333334,
1766
+ "grad_norm": 83.6311264038086,
1767
+ "learning_rate": 1.2700000000000001e-06,
1768
+ "loss": 0.6062012672424316,
1769
+ "margin": 2.137274169921875,
1770
+ "max_reward": 1.9431640625,
1771
+ "mean_reward": -1.1080963134765625,
1772
+ "min_reward": -4.4921875,
1773
+ "num_tokens": 1808403.0,
1774
+ "step": 1120
1775
+ },
1776
+ {
1777
+ "accuracy": 0.75,
1778
+ "epoch": 1.5066666666666668,
1779
+ "grad_norm": 60.31748962402344,
1780
+ "learning_rate": 1.2366666666666668e-06,
1781
+ "loss": 0.7049060821533203,
1782
+ "margin": 1.48701171875,
1783
+ "max_reward": 2.08974609375,
1784
+ "mean_reward": -0.200439453125,
1785
+ "min_reward": -2.463671875,
1786
+ "num_tokens": 1824621.0,
1787
+ "step": 1130
1788
+ },
1789
+ {
1790
+ "accuracy": 0.55,
1791
+ "epoch": 1.52,
1792
+ "grad_norm": 149.60330200195312,
1793
+ "learning_rate": 1.2033333333333334e-06,
1794
+ "loss": 1.1337081909179687,
1795
+ "margin": 0.48056640625,
1796
+ "max_reward": 1.880859375,
1797
+ "mean_reward": -0.384716796875,
1798
+ "min_reward": -2.83515625,
1799
+ "num_tokens": 1840744.0,
1800
+ "step": 1140
1801
+ },
1802
+ {
1803
+ "accuracy": 0.7,
1804
+ "epoch": 1.5333333333333332,
1805
+ "grad_norm": 41.09004592895508,
1806
+ "learning_rate": 1.1700000000000002e-06,
1807
+ "loss": 0.6269871711730957,
1808
+ "margin": 1.2066650390625,
1809
+ "max_reward": 2.41328125,
1810
+ "mean_reward": 0.39190673828125,
1811
+ "min_reward": -1.7814453125,
1812
+ "num_tokens": 1855354.0,
1813
+ "step": 1150
1814
+ },
1815
+ {
1816
+ "epoch": 1.5333333333333332,
1817
+ "eval_accuracy": 0.586,
1818
+ "eval_loss": 0.9035583138465881,
1819
+ "eval_margin": 0.5397208251953125,
1820
+ "eval_max_reward": 1.6990680236816407,
1821
+ "eval_mean_reward": 0.7785431518554687,
1822
+ "eval_min_reward": -0.14198171997070314,
1823
+ "eval_num_tokens": 1855354.0,
1824
+ "eval_runtime": 62.4733,
1825
+ "eval_samples_per_second": 8.003,
1826
+ "eval_steps_per_second": 8.003,
1827
+ "step": 1150
1828
+ },
1829
+ {
1830
+ "accuracy": 0.85,
1831
+ "epoch": 1.5466666666666666,
1832
+ "grad_norm": 6.0896782875061035,
1833
+ "learning_rate": 1.1366666666666667e-06,
1834
+ "loss": 0.5843002796173096,
1835
+ "margin": 1.83350830078125,
1836
+ "max_reward": 2.8265625,
1837
+ "mean_reward": 0.397064208984375,
1838
+ "min_reward": -2.0440673828125,
1839
+ "num_tokens": 1871860.0,
1840
+ "step": 1160
1841
+ },
1842
+ {
1843
+ "accuracy": 0.75,
1844
+ "epoch": 1.56,
1845
+ "grad_norm": 102.23207092285156,
1846
+ "learning_rate": 1.1033333333333335e-06,
1847
+ "loss": 0.5763196468353271,
1848
+ "margin": 1.4167724609375,
1849
+ "max_reward": 3.4875,
1850
+ "mean_reward": 1.43270263671875,
1851
+ "min_reward": -1.278466796875,
1852
+ "num_tokens": 1888601.0,
1853
+ "step": 1170
1854
+ },
1855
+ {
1856
+ "accuracy": 0.6,
1857
+ "epoch": 1.5733333333333333,
1858
+ "grad_norm": 109.7595443725586,
1859
+ "learning_rate": 1.0700000000000001e-06,
1860
+ "loss": 0.9253439903259277,
1861
+ "margin": 0.2265625,
1862
+ "max_reward": 2.1390625,
1863
+ "mean_reward": 0.3791015625,
1864
+ "min_reward": -1.2005859375,
1865
+ "num_tokens": 1905756.0,
1866
+ "step": 1180
1867
+ },
1868
+ {
1869
+ "accuracy": 0.7,
1870
+ "epoch": 1.5866666666666667,
1871
+ "grad_norm": 27.03172492980957,
1872
+ "learning_rate": 1.0366666666666668e-06,
1873
+ "loss": 0.6028451919555664,
1874
+ "margin": 1.306298828125,
1875
+ "max_reward": 2.61015625,
1876
+ "mean_reward": 0.2985595703125,
1877
+ "min_reward": -1.97734375,
1878
+ "num_tokens": 1922025.0,
1879
+ "step": 1190
1880
+ },
1881
+ {
1882
+ "accuracy": 0.75,
1883
+ "epoch": 1.6,
1884
+ "grad_norm": 91.80132293701172,
1885
+ "learning_rate": 1.0033333333333334e-06,
1886
+ "loss": 0.6056195735931397,
1887
+ "margin": 1.401220703125,
1888
+ "max_reward": 2.41943359375,
1889
+ "mean_reward": -0.0202880859375,
1890
+ "min_reward": -2.3765625,
1891
+ "num_tokens": 1938931.0,
1892
+ "step": 1200
1893
+ },
1894
+ {
1895
+ "epoch": 1.6,
1896
+ "eval_accuracy": 0.602,
1897
+ "eval_loss": 1.04457688331604,
1898
+ "eval_margin": 0.740804443359375,
1899
+ "eval_max_reward": 1.606439208984375,
1900
+ "eval_mean_reward": 0.3948948974609375,
1901
+ "eval_min_reward": -0.8166494140625,
1902
+ "eval_num_tokens": 1938931.0,
1903
+ "eval_runtime": 62.3655,
1904
+ "eval_samples_per_second": 8.017,
1905
+ "eval_steps_per_second": 8.017,
1906
+ "step": 1200
1907
+ },
1908
+ {
1909
+ "accuracy": 0.75,
1910
+ "epoch": 1.6133333333333333,
1911
+ "grad_norm": 0.7046879529953003,
1912
+ "learning_rate": 9.7e-07,
1913
+ "loss": 1.0164477348327636,
1914
+ "margin": 1.867919921875,
1915
+ "max_reward": 3.298828125,
1916
+ "mean_reward": 0.1361083984375,
1917
+ "min_reward": -3.41015625,
1918
+ "num_tokens": 1955127.0,
1919
+ "step": 1210
1920
+ },
1921
+ {
1922
+ "accuracy": 0.65,
1923
+ "epoch": 1.6266666666666667,
1924
+ "grad_norm": 25.504459381103516,
1925
+ "learning_rate": 9.366666666666667e-07,
1926
+ "loss": 0.8065451622009278,
1927
+ "margin": 1.255029296875,
1928
+ "max_reward": 2.398828125,
1929
+ "mean_reward": -0.2284912109375,
1930
+ "min_reward": -3.13154296875,
1931
+ "num_tokens": 1972318.0,
1932
+ "step": 1220
1933
+ },
1934
+ {
1935
+ "accuracy": 0.65,
1936
+ "epoch": 1.6400000000000001,
1937
+ "grad_norm": 87.41260528564453,
1938
+ "learning_rate": 9.033333333333334e-07,
1939
+ "loss": 1.119236946105957,
1940
+ "margin": 0.7582611083984375,
1941
+ "max_reward": 2.4771484375,
1942
+ "mean_reward": -0.6033004760742188,
1943
+ "min_reward": -3.24765625,
1944
+ "num_tokens": 1987620.0,
1945
+ "step": 1230
1946
+ },
1947
+ {
1948
+ "accuracy": 0.65,
1949
+ "epoch": 1.6533333333333333,
1950
+ "grad_norm": 94.04547119140625,
1951
+ "learning_rate": 8.7e-07,
1952
+ "loss": 0.6439452171325684,
1953
+ "margin": 1.82333984375,
1954
+ "max_reward": 2.728125,
1955
+ "mean_reward": 0.210400390625,
1956
+ "min_reward": -2.7384765625,
1957
+ "num_tokens": 2004302.0,
1958
+ "step": 1240
1959
+ },
1960
+ {
1961
+ "accuracy": 0.65,
1962
+ "epoch": 1.6666666666666665,
1963
+ "grad_norm": 85.90032958984375,
1964
+ "learning_rate": 8.366666666666668e-07,
1965
+ "loss": 0.5457091331481934,
1966
+ "margin": 1.3715576171875,
1967
+ "max_reward": 2.209033203125,
1968
+ "mean_reward": -0.44652099609375,
1969
+ "min_reward": -2.5546875,
1970
+ "num_tokens": 2021262.0,
1971
+ "step": 1250
1972
+ },
1973
+ {
1974
+ "epoch": 1.6666666666666665,
1975
+ "eval_accuracy": 0.612,
1976
+ "eval_loss": 1.0178015232086182,
1977
+ "eval_margin": 0.7143900146484375,
1978
+ "eval_max_reward": 1.167669921875,
1979
+ "eval_mean_reward": 0.00498980712890625,
1980
+ "eval_min_reward": -1.1576903076171876,
1981
+ "eval_num_tokens": 2021262.0,
1982
+ "eval_runtime": 62.4994,
1983
+ "eval_samples_per_second": 8.0,
1984
+ "eval_steps_per_second": 8.0,
1985
+ "step": 1250
1986
+ },
1987
+ {
1988
+ "accuracy": 0.85,
1989
+ "epoch": 1.6800000000000002,
1990
+ "grad_norm": 57.4267578125,
1991
+ "learning_rate": 8.033333333333335e-07,
1992
+ "loss": 0.5381040573120117,
1993
+ "margin": 1.5513671875,
1994
+ "max_reward": 2.2685546875,
1995
+ "mean_reward": -0.218896484375,
1996
+ "min_reward": -2.88603515625,
1997
+ "num_tokens": 2037311.0,
1998
+ "step": 1260
1999
+ },
2000
+ {
2001
+ "accuracy": 0.65,
2002
+ "epoch": 1.6933333333333334,
2003
+ "grad_norm": 20.574970245361328,
2004
+ "learning_rate": 7.7e-07,
2005
+ "loss": 0.8994636535644531,
2006
+ "margin": 1.5009521484375,
2007
+ "max_reward": 2.207861328125,
2008
+ "mean_reward": -0.48370361328125,
2009
+ "min_reward": -2.67890625,
2010
+ "num_tokens": 2053587.0,
2011
+ "step": 1270
2012
+ },
2013
+ {
2014
+ "accuracy": 0.5,
2015
+ "epoch": 1.7066666666666666,
2016
+ "grad_norm": 25.6247501373291,
2017
+ "learning_rate": 7.366666666666668e-07,
2018
+ "loss": 1.4920889854431152,
2019
+ "margin": 0.5493316650390625,
2020
+ "max_reward": 2.29677734375,
2021
+ "mean_reward": -0.34183807373046876,
2022
+ "min_reward": -3.459832763671875,
2023
+ "num_tokens": 2069715.0,
2024
+ "step": 1280
2025
+ },
2026
+ {
2027
+ "accuracy": 0.65,
2028
+ "epoch": 1.72,
2029
+ "grad_norm": 28.003578186035156,
2030
+ "learning_rate": 7.033333333333334e-07,
2031
+ "loss": 1.0490900039672852,
2032
+ "margin": 1.029248046875,
2033
+ "max_reward": 2.55390625,
2034
+ "mean_reward": -0.0261474609375,
2035
+ "min_reward": -2.5232421875,
2036
+ "num_tokens": 2083729.0,
2037
+ "step": 1290
2038
+ },
2039
+ {
2040
+ "accuracy": 0.6,
2041
+ "epoch": 1.7333333333333334,
2042
+ "grad_norm": 100.13626098632812,
2043
+ "learning_rate": 6.7e-07,
2044
+ "loss": 0.8861888885498047,
2045
+ "margin": 0.58553466796875,
2046
+ "max_reward": 1.53046875,
2047
+ "mean_reward": -0.445611572265625,
2048
+ "min_reward": -2.4734375,
2049
+ "num_tokens": 2099948.0,
2050
+ "step": 1300
2051
+ },
2052
+ {
2053
+ "epoch": 1.7333333333333334,
2054
+ "eval_accuracy": 0.606,
2055
+ "eval_loss": 1.0033907890319824,
2056
+ "eval_margin": 0.7165312805175781,
2057
+ "eval_max_reward": 1.2282371826171874,
2058
+ "eval_mean_reward": 0.07899192810058593,
2059
+ "eval_min_reward": -1.0702533264160157,
2060
+ "eval_num_tokens": 2099948.0,
2061
+ "eval_runtime": 62.3657,
2062
+ "eval_samples_per_second": 8.017,
2063
+ "eval_steps_per_second": 8.017,
2064
+ "step": 1300
2065
+ },
2066
+ {
2067
+ "accuracy": 0.7,
2068
+ "epoch": 1.7466666666666666,
2069
+ "grad_norm": 144.0567626953125,
2070
+ "learning_rate": 6.366666666666667e-07,
2071
+ "loss": 0.5809046268463135,
2072
+ "margin": 1.460107421875,
2073
+ "max_reward": 1.62353515625,
2074
+ "mean_reward": -0.3951416015625,
2075
+ "min_reward": -2.947265625,
2076
+ "num_tokens": 2115789.0,
2077
+ "step": 1310
2078
+ },
2079
+ {
2080
+ "accuracy": 0.7,
2081
+ "epoch": 1.76,
2082
+ "grad_norm": 49.05362319946289,
2083
+ "learning_rate": 6.033333333333334e-07,
2084
+ "loss": 0.87122802734375,
2085
+ "margin": 1.10679931640625,
2086
+ "max_reward": 2.268359375,
2087
+ "mean_reward": -0.050628662109375,
2088
+ "min_reward": -2.1308349609375,
2089
+ "num_tokens": 2133528.0,
2090
+ "step": 1320
2091
+ },
2092
+ {
2093
+ "accuracy": 0.65,
2094
+ "epoch": 1.7733333333333334,
2095
+ "grad_norm": 98.22369384765625,
2096
+ "learning_rate": 5.7e-07,
2097
+ "loss": 0.6749707221984863,
2098
+ "margin": 0.971484375,
2099
+ "max_reward": 2.073046875,
2100
+ "mean_reward": -0.6001953125,
2101
+ "min_reward": -3.078125,
2102
+ "num_tokens": 2147275.0,
2103
+ "step": 1330
2104
+ },
2105
+ {
2106
+ "accuracy": 0.75,
2107
+ "epoch": 1.7866666666666666,
2108
+ "grad_norm": 178.06886291503906,
2109
+ "learning_rate": 5.366666666666667e-07,
2110
+ "loss": 0.7042564868927002,
2111
+ "margin": 1.1953125,
2112
+ "max_reward": 1.126953125,
2113
+ "mean_reward": -1.3919921875,
2114
+ "min_reward": -3.98046875,
2115
+ "num_tokens": 2165353.0,
2116
+ "step": 1340
2117
+ },
2118
+ {
2119
+ "accuracy": 0.65,
2120
+ "epoch": 1.8,
2121
+ "grad_norm": 49.11288070678711,
2122
+ "learning_rate": 5.033333333333334e-07,
2123
+ "loss": 0.9773781776428223,
2124
+ "margin": 0.57353515625,
2125
+ "max_reward": 1.633447265625,
2126
+ "mean_reward": -0.50439453125,
2127
+ "min_reward": -2.768359375,
2128
+ "num_tokens": 2181126.0,
2129
+ "step": 1350
2130
+ },
2131
+ {
2132
+ "epoch": 1.8,
2133
+ "eval_accuracy": 0.612,
2134
+ "eval_loss": 0.9391070008277893,
2135
+ "eval_margin": 0.6821452369689941,
2136
+ "eval_max_reward": 0.2567112731933594,
2137
+ "eval_mean_reward": -0.8024592456817627,
2138
+ "eval_min_reward": -1.8616297645568847,
2139
+ "eval_num_tokens": 2181126.0,
2140
+ "eval_runtime": 62.4622,
2141
+ "eval_samples_per_second": 8.005,
2142
+ "eval_steps_per_second": 8.005,
2143
+ "step": 1350
2144
+ },
2145
+ {
2146
+ "accuracy": 0.8,
2147
+ "epoch": 1.8133333333333335,
2148
+ "grad_norm": 12.369355201721191,
2149
+ "learning_rate": 4.7000000000000005e-07,
2150
+ "loss": 0.5082510471343994,
2151
+ "margin": 2.278076171875,
2152
+ "max_reward": 2.44150390625,
2153
+ "mean_reward": -0.1749755859375,
2154
+ "min_reward": -3.0359375,
2155
+ "num_tokens": 2198470.0,
2156
+ "step": 1360
2157
+ },
2158
+ {
2159
+ "accuracy": 0.6,
2160
+ "epoch": 1.8266666666666667,
2161
+ "grad_norm": 21.943758010864258,
2162
+ "learning_rate": 4.366666666666667e-07,
2163
+ "loss": 0.8553391456604004,
2164
+ "margin": 1.10614013671875,
2165
+ "max_reward": 0.610546875,
2166
+ "mean_reward": -1.453851318359375,
2167
+ "min_reward": -3.56328125,
2168
+ "num_tokens": 2213797.0,
2169
+ "step": 1370
2170
+ },
2171
+ {
2172
+ "accuracy": 0.55,
2173
+ "epoch": 1.8399999999999999,
2174
+ "grad_norm": 67.34660339355469,
2175
+ "learning_rate": 4.0333333333333337e-07,
2176
+ "loss": 1.136549472808838,
2177
+ "margin": 0.629931640625,
2178
+ "max_reward": 2.42353515625,
2179
+ "mean_reward": -0.6287841796875,
2180
+ "min_reward": -3.7546875,
2181
+ "num_tokens": 2230299.0,
2182
+ "step": 1380
2183
+ },
2184
+ {
2185
+ "accuracy": 0.65,
2186
+ "epoch": 1.8533333333333335,
2187
+ "grad_norm": 14.513836860656738,
2188
+ "learning_rate": 3.7e-07,
2189
+ "loss": 0.5644237518310546,
2190
+ "margin": 1.4962158203125,
2191
+ "max_reward": 1.477099609375,
2192
+ "mean_reward": -1.04271240234375,
2193
+ "min_reward": -3.7123046875,
2194
+ "num_tokens": 2245618.0,
2195
+ "step": 1390
2196
+ },
2197
+ {
2198
+ "accuracy": 0.5,
2199
+ "epoch": 1.8666666666666667,
2200
+ "grad_norm": 137.32266235351562,
2201
+ "learning_rate": 3.366666666666667e-07,
2202
+ "loss": 0.7403782367706299,
2203
+ "margin": 1.222882080078125,
2204
+ "max_reward": 1.35396728515625,
2205
+ "mean_reward": -1.0289886474609375,
2206
+ "min_reward": -3.173046875,
2207
+ "num_tokens": 2262798.0,
2208
+ "step": 1400
2209
+ },
2210
+ {
2211
+ "epoch": 1.8666666666666667,
2212
+ "eval_accuracy": 0.618,
2213
+ "eval_loss": 0.9525383710861206,
2214
+ "eval_margin": 0.7151371459960938,
2215
+ "eval_max_reward": -0.032703857421875,
2216
+ "eval_mean_reward": -1.126583221435547,
2217
+ "eval_min_reward": -2.220462585449219,
2218
+ "eval_num_tokens": 2262798.0,
2219
+ "eval_runtime": 62.453,
2220
+ "eval_samples_per_second": 8.006,
2221
+ "eval_steps_per_second": 8.006,
2222
+ "step": 1400
2223
+ },
2224
+ {
2225
+ "accuracy": 0.75,
2226
+ "epoch": 1.88,
2227
+ "grad_norm": 87.93866729736328,
2228
+ "learning_rate": 3.033333333333334e-07,
2229
+ "loss": 0.7155962944030761,
2230
+ "margin": 1.211572265625,
2231
+ "max_reward": 1.64248046875,
2232
+ "mean_reward": -0.8534912109375,
2233
+ "min_reward": -3.3265625,
2234
+ "num_tokens": 2280625.0,
2235
+ "step": 1410
2236
+ },
2237
+ {
2238
+ "accuracy": 0.6,
2239
+ "epoch": 1.8933333333333333,
2240
+ "grad_norm": 136.62033081054688,
2241
+ "learning_rate": 2.7e-07,
2242
+ "loss": 0.7803818702697753,
2243
+ "margin": 1.5741943359375,
2244
+ "max_reward": 1.125,
2245
+ "mean_reward": -1.46397705078125,
2246
+ "min_reward": -3.69921875,
2247
+ "num_tokens": 2297438.0,
2248
+ "step": 1420
2249
+ },
2250
+ {
2251
+ "accuracy": 0.8,
2252
+ "epoch": 1.9066666666666667,
2253
+ "grad_norm": 36.08385467529297,
2254
+ "learning_rate": 2.3666666666666667e-07,
2255
+ "loss": 0.44714956283569335,
2256
+ "margin": 1.790771484375,
2257
+ "max_reward": 2.1365234375,
2258
+ "mean_reward": -1.1315673828125,
2259
+ "min_reward": -4.0796875,
2260
+ "num_tokens": 2315559.0,
2261
+ "step": 1430
2262
+ },
2263
+ {
2264
+ "accuracy": 0.75,
2265
+ "epoch": 1.92,
2266
+ "grad_norm": 16.93815803527832,
2267
+ "learning_rate": 2.0333333333333333e-07,
2268
+ "loss": 0.6091221332550049,
2269
+ "margin": 1.25640869140625,
2270
+ "max_reward": 0.8255126953125,
2271
+ "mean_reward": -1.083807373046875,
2272
+ "min_reward": -3.1236328125,
2273
+ "num_tokens": 2331603.0,
2274
+ "step": 1440
2275
+ },
2276
+ {
2277
+ "accuracy": 0.55,
2278
+ "epoch": 1.9333333333333333,
2279
+ "grad_norm": 53.92049026489258,
2280
+ "learning_rate": 1.7000000000000001e-07,
2281
+ "loss": 0.7587657928466797,
2282
+ "margin": 1.339501953125,
2283
+ "max_reward": 1.06416015625,
2284
+ "mean_reward": -0.8656494140625,
2285
+ "min_reward": -3.60859375,
2286
+ "num_tokens": 2346680.0,
2287
+ "step": 1450
2288
+ },
2289
+ {
2290
+ "epoch": 1.9333333333333333,
2291
+ "eval_accuracy": 0.612,
2292
+ "eval_loss": 0.97000652551651,
2293
+ "eval_margin": 0.7449773254394532,
2294
+ "eval_max_reward": 0.27565774536132814,
2295
+ "eval_mean_reward": -0.8554528350830078,
2296
+ "eval_min_reward": -1.9865634155273437,
2297
+ "eval_num_tokens": 2346680.0,
2298
+ "eval_runtime": 62.49,
2299
+ "eval_samples_per_second": 8.001,
2300
+ "eval_steps_per_second": 8.001,
2301
+ "step": 1450
2302
+ },
2303
+ {
2304
+ "accuracy": 0.8,
2305
+ "epoch": 1.9466666666666668,
2306
+ "grad_norm": 59.9929084777832,
2307
+ "learning_rate": 1.3666666666666667e-07,
2308
+ "loss": 0.46772356033325196,
2309
+ "margin": 1.48759765625,
2310
+ "max_reward": 1.7115234375,
2311
+ "mean_reward": -0.782666015625,
2312
+ "min_reward": -2.921875,
2313
+ "num_tokens": 2363335.0,
2314
+ "step": 1460
2315
+ },
2316
+ {
2317
+ "accuracy": 0.85,
2318
+ "epoch": 1.96,
2319
+ "grad_norm": 67.4390869140625,
2320
+ "learning_rate": 1.0333333333333335e-07,
2321
+ "loss": 0.37448573112487793,
2322
+ "margin": 1.778466796875,
2323
+ "max_reward": 0.9283203125,
2324
+ "mean_reward": -1.5448974609375,
2325
+ "min_reward": -3.9453125,
2326
+ "num_tokens": 2379720.0,
2327
+ "step": 1470
2328
+ },
2329
+ {
2330
+ "accuracy": 0.75,
2331
+ "epoch": 1.9733333333333334,
2332
+ "grad_norm": 93.11763763427734,
2333
+ "learning_rate": 7e-08,
2334
+ "loss": 0.4680135726928711,
2335
+ "margin": 1.938623046875,
2336
+ "max_reward": 2.512109375,
2337
+ "mean_reward": 0.1829833984375,
2338
+ "min_reward": -2.870703125,
2339
+ "num_tokens": 2394956.0,
2340
+ "step": 1480
2341
+ },
2342
+ {
2343
+ "accuracy": 0.85,
2344
+ "epoch": 1.9866666666666668,
2345
+ "grad_norm": 40.56834030151367,
2346
+ "learning_rate": 3.666666666666667e-08,
2347
+ "loss": 0.35323076248168944,
2348
+ "margin": 1.9112945556640626,
2349
+ "max_reward": 1.086480712890625,
2350
+ "mean_reward": -1.4766769409179688,
2351
+ "min_reward": -3.66484375,
2352
+ "num_tokens": 2410974.0,
2353
+ "step": 1490
2354
+ },
2355
+ {
2356
+ "accuracy": 0.65,
2357
+ "epoch": 2.0,
2358
+ "grad_norm": 85.22586822509766,
2359
+ "learning_rate": 3.3333333333333334e-09,
2360
+ "loss": 0.5986483097076416,
2361
+ "margin": 1.537744140625,
2362
+ "max_reward": 2.01474609375,
2363
+ "mean_reward": -0.6834228515625,
2364
+ "min_reward": -3.82421875,
2365
+ "num_tokens": 2425614.0,
2366
+ "step": 1500
2367
+ },
2368
+ {
2369
+ "epoch": 2.0,
2370
+ "eval_accuracy": 0.616,
2371
+ "eval_loss": 0.978255033493042,
2372
+ "eval_margin": 0.7540488891601562,
2373
+ "eval_max_reward": 0.32203082275390627,
2374
+ "eval_mean_reward": -0.8245922546386719,
2375
+ "eval_min_reward": -1.97121533203125,
2376
+ "eval_num_tokens": 2425614.0,
2377
+ "eval_runtime": 62.4138,
2378
+ "eval_samples_per_second": 8.011,
2379
+ "eval_steps_per_second": 8.011,
2380
+ "step": 1500
2381
+ }
2382
+ ],
2383
+ "logging_steps": 10,
2384
+ "max_steps": 1500,
2385
+ "num_input_tokens_seen": 0,
2386
+ "num_train_epochs": 2,
2387
+ "save_steps": 50,
2388
+ "stateful_callbacks": {
2389
+ "TrainerControl": {
2390
+ "args": {
2391
+ "should_epoch_stop": false,
2392
+ "should_evaluate": false,
2393
+ "should_log": false,
2394
+ "should_save": true,
2395
+ "should_training_stop": true
2396
+ },
2397
+ "attributes": {}
2398
+ }
2399
+ },
2400
+ "total_flos": 2054300643182592.0,
2401
+ "train_batch_size": 2,
2402
+ "trial_name": null,
2403
+ "trial_params": null
2404
+ }
checkpoint-1500/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d5bcd49bd6d1025adc5d4385da3c333196b4008baf24052fee3d16f53db65152
3
+ size 5393
config.json ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "activation_function": "gelu_pytorch_tanh",
3
+ "add_cross_attention": false,
4
+ "architectures": [
5
+ "GPTBigCodeForSequenceClassification"
6
+ ],
7
+ "attention_softmax_in_fp32": true,
8
+ "attn_pdrop": 0.1,
9
+ "bos_token_id": 0,
10
+ "dtype": "float32",
11
+ "embd_pdrop": 0.1,
12
+ "eos_token_id": 0,
13
+ "id2label": {
14
+ "0": "LABEL_0"
15
+ },
16
+ "inference_runner": 0,
17
+ "initializer_range": 0.02,
18
+ "label2id": {
19
+ "LABEL_0": 0
20
+ },
21
+ "layer_norm_epsilon": 1e-05,
22
+ "max_batch_size": null,
23
+ "max_sequence_length": null,
24
+ "model_type": "gpt_bigcode",
25
+ "multi_query": true,
26
+ "n_embd": 768,
27
+ "n_head": 12,
28
+ "n_inner": 3072,
29
+ "n_layer": 20,
30
+ "n_positions": 8192,
31
+ "num_key_value_heads": 1,
32
+ "pad_key_length": true,
33
+ "pad_token_id": 0,
34
+ "pre_allocate_kv_cache": false,
35
+ "resid_pdrop": 0.1,
36
+ "scale_attention_softmax_in_fp32": true,
37
+ "scale_attn_weights": true,
38
+ "summary_activation": null,
39
+ "summary_first_dropout": 0.1,
40
+ "summary_proj_to_labels": true,
41
+ "summary_type": "cls_index",
42
+ "summary_use_proj": true,
43
+ "tie_word_embeddings": true,
44
+ "transformers_version": "5.0.0",
45
+ "use_cache": false,
46
+ "validate_runner_input": true,
47
+ "vocab_size": 49152
48
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:700a6584a30ad5bbfe191ea3c16f9cc5f1c1f94c65d782fa167b907d55fbe920
3
+ size 656604448
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "backend": "tokenizers",
4
+ "bos_token": "<|endoftext|>",
5
+ "eos_token": "<|endoftext|>",
6
+ "errors": "replace",
7
+ "extra_special_tokens": [
8
+ "<|endoftext|>",
9
+ "<fim_prefix>",
10
+ "<fim_middle>",
11
+ "<fim_suffix>",
12
+ "<fim_pad>",
13
+ "<filename>",
14
+ "<gh_stars>",
15
+ "<issue_start>",
16
+ "<issue_comment>",
17
+ "<issue_closed>",
18
+ "<jupyter_start>",
19
+ "<jupyter_text>",
20
+ "<jupyter_code>",
21
+ "<jupyter_output>",
22
+ "<empty_output>",
23
+ "<commit_before>",
24
+ "<commit_msg>",
25
+ "<commit_after>",
26
+ "<reponame>"
27
+ ],
28
+ "is_local": false,
29
+ "model_max_length": 1000000000000000019884624838656,
30
+ "pad_token": "<|endoftext|>",
31
+ "tokenizer_class": "GPT2Tokenizer",
32
+ "unk_token": "<|endoftext|>",
33
+ "vocab_size": 49152
34
+ }
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d5bcd49bd6d1025adc5d4385da3c333196b4008baf24052fee3d16f53db65152
3
+ size 5393