Transformers
Safetensors
Laya
system-1
action-ranking
decision-model
enterprise-ai
research
experimental
Instructions to use yasserrmd/enterprise-reflux-laya-v21 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use yasserrmd/enterprise-reflux-laya-v21 with Transformers:
# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("yasserrmd/enterprise-reflux-laya-v21", device_map="auto") - Laya
How to use yasserrmd/enterprise-reflux-laya-v21 with Laya:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
Enterprise Reflux Laya V2.1: targeted state/NO_ACTION + ACT/ESCALATE training
Browse files- README.md +58 -0
- act_recovery_config.json +6 -0
- encoder/config.json +84 -0
- enterprise_reflux_policy.json +8 -0
- experiment_metadata.json +17 -0
- model.safetensors +3 -0
- rl_agent_config.json +43 -0
- tokenizer/tokenizer.json +0 -0
- tokenizer/tokenizer_config.json +15 -0
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 |
+
}
|