yasserrmd commited on
Commit
8985516
·
verified ·
1 Parent(s): 656430c

Enterprise Reflux Laya V2.1: targeted state/NO_ACTION + ACT/ESCALATE training

Browse files
README.md ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ library_name: transformers
4
+ base_model: convaiinnovations/laya
5
+ tags:
6
+ - laya
7
+ - system-1
8
+ - decision-model
9
+ - enterprise
10
+ - action-ranking
11
+ - act-escalate
12
+ ---
13
+
14
+ # Enterprise Reflux Laya V2.1
15
+
16
+ Targeted specialization of the public Laya decision model for enterprise
17
+ dynamic action selection.
18
+
19
+ ## What changed from the first Enterprise Reflux Laya run
20
+
21
+ - Added targeted V2.1 training data:
22
+ - counterfactual state flips
23
+ - hard NO_ACTION
24
+ - sibling actions
25
+ - policy constraints
26
+ - ACT / ESCALATE supervision
27
+ - cross-domain collisions
28
+ - Preserved the official Laya architecture.
29
+ - Preserved RLCD + soft-CE choice training.
30
+ - Added supervised ACT/ESCALATE cross-entropy.
31
+ - Reinitialized the shipped ACT head before training because the original
32
+ `act_probability` head is known to saturate.
33
+ - Class 0 is ACT because the Laya runtime publishes
34
+ `softmax(act_logits)[0]` as `act_probability`.
35
+ - Fitted choice temperature on held-out calibration data.
36
+ - Fitted ACT temperature on held-out V2.1 calibration data and baked it
37
+ into the final ACT linear layer.
38
+ - Selected the System-1 ACT threshold from calibration data only, with a
39
+ minimum handled-sample requirement.
40
+
41
+ ## System-1 policy
42
+
43
+ ACT threshold: `0.6400`
44
+
45
+ Threshold selection status: `target_not_met`
46
+
47
+ Choice temperature: `3.1137`
48
+
49
+ ACT temperature baked into model: `1.1130`
50
+
51
+ The neural model is not an authorization engine. Deterministic eligibility,
52
+ authorization, policy enforcement, approval checks and irreversible-action
53
+ controls remain external.
54
+
55
+ ## Benchmark integrity
56
+
57
+ The frozen-200, hard-400 and mixed-500 evaluation sets were not used for
58
+ training, calibration, or threshold fitting.
act_recovery_config.json ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ {
2
+ "apply_layer_norm_before_act_head": true,
3
+ "normalized_shape": "dynamic_last_dimension",
4
+ "act_class": 0,
5
+ "escalate_class": 1
6
+ }
encoder/config.json ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "ModernBertForMaskedLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 50281,
8
+ "classifier_activation": "gelu",
9
+ "classifier_bias": false,
10
+ "classifier_dropout": 0.0,
11
+ "classifier_pooling": "mean",
12
+ "cls_token_id": 50281,
13
+ "decoder_bias": true,
14
+ "deterministic_flash_attn": false,
15
+ "dtype": "float32",
16
+ "embedding_dropout": 0.0,
17
+ "eos_token_id": 50282,
18
+ "global_attn_every_n_layers": 3,
19
+ "gradient_checkpointing": false,
20
+ "hidden_activation": "gelu",
21
+ "hidden_size": 1024,
22
+ "initializer_cutoff_factor": 2.0,
23
+ "initializer_range": 0.02,
24
+ "intermediate_size": 2624,
25
+ "layer_norm_eps": 1e-05,
26
+ "layer_types": [
27
+ "full_attention",
28
+ "sliding_attention",
29
+ "sliding_attention",
30
+ "full_attention",
31
+ "sliding_attention",
32
+ "sliding_attention",
33
+ "full_attention",
34
+ "sliding_attention",
35
+ "sliding_attention",
36
+ "full_attention",
37
+ "sliding_attention",
38
+ "sliding_attention",
39
+ "full_attention",
40
+ "sliding_attention",
41
+ "sliding_attention",
42
+ "full_attention",
43
+ "sliding_attention",
44
+ "sliding_attention",
45
+ "full_attention",
46
+ "sliding_attention",
47
+ "sliding_attention",
48
+ "full_attention",
49
+ "sliding_attention",
50
+ "sliding_attention",
51
+ "full_attention",
52
+ "sliding_attention",
53
+ "sliding_attention",
54
+ "full_attention"
55
+ ],
56
+ "local_attention": 128,
57
+ "max_position_embeddings": 8192,
58
+ "mlp_bias": false,
59
+ "mlp_dropout": 0.0,
60
+ "model_type": "modernbert",
61
+ "norm_bias": false,
62
+ "norm_eps": 1e-05,
63
+ "num_attention_heads": 16,
64
+ "num_hidden_layers": 28,
65
+ "pad_token_id": 50283,
66
+ "position_embedding_type": "absolute",
67
+ "repad_logits_with_grad": false,
68
+ "rope_parameters": {
69
+ "full_attention": {
70
+ "rope_theta": 160000.0,
71
+ "rope_type": "default"
72
+ },
73
+ "sliding_attention": {
74
+ "rope_theta": 10000.0,
75
+ "rope_type": "default"
76
+ }
77
+ },
78
+ "sep_token_id": 50282,
79
+ "sparse_pred_ignore_index": -100,
80
+ "sparse_prediction": false,
81
+ "tie_word_embeddings": true,
82
+ "transformers_version": "5.16.1",
83
+ "vocab_size": 50368
84
+ }
enterprise_reflux_policy.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "system1_act_threshold": 0.6400000000000001,
3
+ "choice_temperature": 3.1136913299560547,
4
+ "act_temperature_baked": 1.1129868030548096,
5
+ "act_class": 0,
6
+ "escalate_class": 1,
7
+ "note": "Deterministic authorization and hard policy eligibility remain external to the neural model."
8
+ }
experiment_metadata.json ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "V2.1 trained ranking model + recovered ACT head",
3
+ "act_recovery": true,
4
+ "act_validation_accuracy": 0.606947,
5
+ "act_validation_balanced_accuracy": 0.610865,
6
+ "act_validation_auroc": 0.6769,
7
+ "stage1_gate_auroc": 0.8646,
8
+ "stage2_gate_selected": "logistic",
9
+ "stage2_gate_val_selective_accuracy": 0.7981,
10
+ "stage2_gate_val_coverage": 0.0951,
11
+ "notes": [
12
+ "Ranking model preserved",
13
+ "ACT head recovered from collapsed state",
14
+ "Execution gate not suitable yet for autonomous System-1 routing",
15
+ "Authorization/policy/preconditions remain external"
16
+ ]
17
+ }
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:920ede8882341d082701886330c58bc393cad0ab51afc0c992164060e80383c4
3
+ size 842609220
rl_agent_config.json ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "encoder": "answerdotai/ModernBERT-large",
3
+ "head_layers": 2,
4
+ "max_len": 1024,
5
+ "head_max_len": 256,
6
+ "max_prefixes": 6,
7
+ "act_costs": {
8
+ "escalate": 0.5
9
+ },
10
+ "cost_wrong_act": 3.0,
11
+ "amp_dtype": "bf16",
12
+ "model_name": "enterprise-reflux-laya-v21",
13
+ "temperature": [
14
+ 3.1136913299560547,
15
+ 1.0,
16
+ 1.0
17
+ ],
18
+ "temperature_by_options": {},
19
+ "training": {
20
+ "recipe": "Laya RLCD + soft CE + supervised ACT/ESCALATE",
21
+ "base_model": "convaiinnovations/laya",
22
+ "v21_dataset": "yasserrmd/enterprise-reflux-laya-v21-dataset",
23
+ "epochs": 4,
24
+ "updates": 4284,
25
+ "encoder_lr": 2.5e-05,
26
+ "head_lr": 0.0001,
27
+ "group_size": 4,
28
+ "sigma_start": 0.4,
29
+ "sigma_end": 0.1,
30
+ "act_loss_weight": 0.25,
31
+ "act_class_zero": "ACT",
32
+ "act_class_one": "ESCALATE",
33
+ "act_temperature_baked": 1.1129868030548096,
34
+ "system1_act_threshold": 0.6400000000000001,
35
+ "threshold_status": "target_not_met",
36
+ "minimum_calibration_handled": 56,
37
+ "calibration_held_out_before_training": true,
38
+ "frozen_benchmarks_used_for_training": false
39
+ },
40
+ "gradient_checkpointing": true,
41
+ "max_tokens_per_batch": 4096,
42
+ "fine_tuned": true
43
+ }
tokenizer/tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer/tokenizer_config.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "clean_up_tokenization_spaces": true,
3
+ "cls_token": "[CLS]",
4
+ "local_files_only": false,
5
+ "mask_token": "[MASK]",
6
+ "model_input_names": [
7
+ "input_ids",
8
+ "attention_mask"
9
+ ],
10
+ "model_max_length": 8192,
11
+ "pad_token": "[PAD]",
12
+ "sep_token": "[SEP]",
13
+ "tokenizer_class": "PreTrainedTokenizerFast",
14
+ "unk_token": "[UNK]"
15
+ }