Publish accepted BehaviorTune V1.1-R1 QLoRA adapter

PROVENANCE_AND_HASHES.json

@@ -0,0 +1,63 @@
+{
+  "schema_version": 1,
+  "artifact": "BehaviorTune V1.1-R1 public release provenance",
+  "status": "PREPARED_PENDING_PUBLICATION",
+  "source_commit": "6ef019e996982417df50127a5c351f179b75ea6b",
+  "release_tag": "v1.0.0",
+  "public_targets": {
+    "github": "https://github.com/aamish-ahmad/behavior-tune",
+    "dataset": "https://huggingface.co/datasets/aamish-ahmad/behaviortune-v1-1-r1",
+    "adapter": "https://huggingface.co/aamish-ahmad/behaviortune-v1-1-r1-adapter"
+  },
+  "license": {
+    "identifier": "other",
+    "grant": "none",
+    "file": "LICENSE"
+  },
+  "base_model": {
+    "id": "Qwen/Qwen3-4B-Instruct-2507",
+    "revision": "cdbee75f17c01a7cc42f958dc650907174af0554"
+  },
+  "dataset": {
+    "version": "V1.1-R1-1",
+    "scenario_count": 544,
+    "pair_count": 272,
+    "manifest_sha256": "3d961add3fe6dd6737d1fea6ffd6bdb9a56d9d21f7551b66a25e8424f36f5a08",
+    "freeze_manifest_sha256": "84502f301a56ef41551c267ea2445e6bb07d136bdcf2868cee4fa5a4621ae199",
+    "splits": {
+      "train.jsonl": "be16a7c198cacd727992f4ff82493ec40e5c042f9a485e2091a62eab213f91eb",
+      "dev.jsonl": "5eed447ba0878e4a6e26729476ef43d5bd59966469558fc1a0545af904023e5f",
+      "eval_core.jsonl": "807eb5c7ef3bede23aa2fbf44fa4ad3607877835256cdad9383518953c681c1d",
+      "holdout_principal.jsonl": "48fef33ebfb7a8d4f3b46ad4b8c3eb6befcb4c542d6af828455c5860e8e344fc",
+      "holdout_family.jsonl": "f823607a3f42b2c549c502a9958fb805a75beff18fdf1d177d839f752a166df2",
+      "holdout_joint.jsonl": "0cc76df04c5cc394b4f9f46938c6d2605328caffbb9930b8907494f3bb44cd24"
+    }
+  },
+  "adapter": {
+    "recipe_id": "behaviortune-v1-qlora-primary-r1",
+    "adapter_config_sha256": "d6ebc4b0b2ada4e0020aef4f83d95f7eb4745defe9c4ecacf6d1f3aaca11111f",
+    "adapter_model_sha256": "8d16ef2cb6ff7a982511fd58f21eff52538761f4d198b4cc5cbfd73ca7c9d4de",
+    "training_manifest_sha256": "a3115ea2df34452874a860d9857c01ee413b7fa1051edf7889f5668e3186fc62",
+    "retry_executed": false
+  },
+  "accepted_evidence": {
+    "sensitivity_final_verification_sha256": "93e15eb660417d38dff73ee7808659ba7b20f5763271d68b204ad3035df1ebef",
+    "qlora_final_verification_sha256": "07e3d5656459ef11837107cac99e98c371cce08c4c9278ef267de868cacc929a",
+    "metrics_sha256": "4396b6cc47383c974f0fa17fc3e25325fc3a4a2418d7557530fcffe2164db98e",
+    "gate_decision_sha256": "b4096f953dcce5c832549013c1e21245c2c02836c91ad03453989d287723dfd0",
+    "g9_verification_sha256": "b8dd40e9e7639ee0ae9d6911b06182bbb91c1fcec839bf2f2f6a1406fe6c866a"
+  },
+  "result": {
+    "qlora_minus_base_activation_delta": 0.34375,
+    "paired_bootstrap_95": [0.1875, 0.5],
+    "frozen_gates_passed": 6,
+    "frozen_gates_total": 6
+  },
+  "forbidden_work_not_performed": [
+    "new model training",
+    "new inference",
+    "benchmark redesign",
+    "holdout observation",
+    "scientific result mutation"
+  ]
+}

README.md

@@ -0,0 +1,64 @@
+---
+base_model: Qwen/Qwen3-4B-Instruct-2507
+library_name: peft
+pipeline_tag: text-generation
+license: other
+datasets:
+- aamish-ahmad/behaviortune-v1-1-r1
+tags:
+- peft
+- lora
+- qlora
+- behavioral-evaluation
+---
+
+# BehaviorTune V1.1-R1 QLoRA adapter
+
+The accepted BehaviorTune adapter for principal-conditioned binary-choice
+behavior. It was produced by one bounded completion-only SFT/PEFT QLoRA run.
+
+## Identity
+
+- Base model: `Qwen/Qwen3-4B-Instruct-2507`
+- Required base revision: `cdbee75f17c01a7cc42f958dc650907174af0554`
+- Adapter SHA-256: `8d16ef2cb6ff7a982511fd58f21eff52538761f4d198b4cc5cbfd73ca7c9d4de`
+- Dataset: [BehaviorTune Trajectories V1.1-R1](https://huggingface.co/datasets/aamish-ahmad/behaviortune-v1-1-r1)
+- Code and evidence: [GitHub release](https://github.com/aamish-ahmad/behavior-tune/tree/v1.0.0)
+
+## Training
+
+| Item | Value |
+| --- | --- |
+| Rows | 240 |
+| Epochs | 3 |
+| Optimizer steps | 90 |
+| Seed / data seed | 147 / 147 |
+| Quantization | 4-bit NF4, BF16 compute |
+| LoRA | rank 32, alpha 64, dropout 0.05, all linear projections |
+| Effective batch size | 8 |
+| Learning rate | 0.0002, linear schedule |
+| Hardware | NVIDIA A100-SXM4-40GB |
+
+Exact dependencies and hyperparameters are in the public [training manifest](https://github.com/aamish-ahmad/behavior-tune/blob/v1.0.0/artifacts/behaviortune-v11-r1-qlora-core-20260904/training_evidence/training_manifest.json).
+
+## Evaluation
+
+The matched evaluation used 64 frozen `eval_core` scenarios for each of BASE,
+SYSTEM, CONTEXT, and QLoRA (256 outputs total). QLoRA format-valid activation
+was `1.0` versus BASE `0.65625`, a matched delta of `0.34375`. The predeclared
+paired-bootstrap 95% interval was `[0.1875, 0.5]`; all six frozen gates passed.
+Specificity was `1.0`, false-favor rate was `0.0`, and format-failure rate was
+`0.0` for QLoRA. No retry or second scientific run was executed.
+
+See the public [results](https://github.com/aamish-ahmad/behavior-tune/blob/v1.0.0/docs/RESULTS.md), [metrics](https://github.com/aamish-ahmad/behavior-tune/blob/v1.0.0/artifacts/behaviortune-v11-r1-qlora-core-20260904/evaluation_evidence/metrics.json), [gate decision](https://github.com/aamish-ahmad/behavior-tune/blob/v1.0.0/artifacts/behaviortune-v11-r1-qlora-core-20260904/evaluation_evidence/gate_decision.json), and [independent verification](https://github.com/aamish-ahmad/behavior-tune/blob/v1.0.0/artifacts/behaviortune-v11-r1-qlora-core-20260904/FINAL_VERIFICATION.json).
+
+## Use and limitations
+
+Load the pinned base revision, then attach this PEFT adapter. It is a research
+and portfolio proof object, not a general decision system. The accepted R1 run
+did not evaluate holdouts, LONG-NEUTRAL, persistence, or remediation. Do not
+generalize the observed `eval_core` effect beyond that boundary.
+
+`license: other` denotes that no open-source model license is granted for this
+adapter. The upstream base model remains governed by its own license. Contact
+the copyright holder for permission beyond inspection and evaluation.

adapter_config.json

@@ -0,0 +1,39 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-4B-Instruct-2507",
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 32,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "q_proj",
+    "o_proj",
+    "up_proj",
+    "k_proj",
+    "gate_proj",
+    "v_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d16ef2cb6ff7a982511fd58f21eff52538761f4d198b4cc5cbfd73ca7c9d4de
+size 264308896

evaluation_metrics.json

@@ -0,0 +1,387 @@
+{
+  "condition_rates_and_wilson_95": {
+    "BASE": {
+      "AR_all": {
+        "successes": 21,
+        "total": 32,
+        "value": 0.65625,
+        "wilson_95": {
+          "lower": 0.4831108797896222,
+          "upper": 0.7958956207556682
+        }
+      },
+      "AR_valid": {
+        "successes": 21,
+        "total": 32,
+        "value": 0.65625,
+        "wilson_95": {
+          "lower": 0.4831108797896222,
+          "upper": 0.7958956207556682
+        }
+      },
+      "BA_all": {
+        "successes": 53,
+        "total": 64,
+        "value": 0.828125,
+        "wilson_95": {
+          "lower": 0.7178678837734179,
+          "upper": 0.9012225769875536
+        }
+      },
+      "BA_valid": {
+        "successes": 53,
+        "total": 64,
+        "value": 0.828125,
+        "wilson_95": {
+          "lower": 0.7178678837734179,
+          "upper": 0.9012225769875536
+        }
+      },
+      "EFR_C_valid": {
+        "successes": 16,
+        "total": 16,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8063923194655637,
+          "upper": 1.0000000000000002
+        }
+      },
+      "EFR_D_valid": {
+        "successes": 16,
+        "total": 16,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8063923194655637,
+          "upper": 1.0000000000000002
+        }
+      },
+      "FAR_valid": {
+        "successes": 0,
+        "total": 16,
+        "value": 0.0,
+        "wilson_95": {
+          "lower": 0.0,
+          "upper": 0.1936076805344365
+        }
+      },
+      "FFR": {
+        "successes": 0,
+        "total": 64,
+        "value": 0.0,
+        "wilson_95": {
+          "lower": 0.0,
+          "upper": 0.05662405979280534
+        }
+      },
+      "SPEC_all": {
+        "successes": 32,
+        "total": 32,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8928208017449293,
+          "upper": 1.0
+        }
+      },
+      "SPEC_valid": {
+        "successes": 32,
+        "total": 32,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8928208017449293,
+          "upper": 1.0
+        }
+      }
+    },
+    "CONTEXT": {
+      "AR_all": {
+        "successes": 32,
+        "total": 32,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8928208017449293,
+          "upper": 1.0
+        }
+      },
+      "AR_valid": {
+        "successes": 32,
+        "total": 32,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8928208017449293,
+          "upper": 1.0
+        }
+      },
+      "BA_all": {
+        "successes": 64,
+        "total": 64,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.9433759402071946,
+          "upper": 1.0
+        }
+      },
+      "BA_valid": {
+        "successes": 64,
+        "total": 64,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.9433759402071946,
+          "upper": 1.0
+        }
+      },
+      "EFR_C_valid": {
+        "successes": 16,
+        "total": 16,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8063923194655637,
+          "upper": 1.0000000000000002
+        }
+      },
+      "EFR_D_valid": {
+        "successes": 16,
+        "total": 16,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8063923194655637,
+          "upper": 1.0000000000000002
+        }
+      },
+      "FAR_valid": {
+        "successes": 0,
+        "total": 16,
+        "value": 0.0,
+        "wilson_95": {
+          "lower": 0.0,
+          "upper": 0.1936076805344365
+        }
+      },
+      "FFR": {
+        "successes": 0,
+        "total": 64,
+        "value": 0.0,
+        "wilson_95": {
+          "lower": 0.0,
+          "upper": 0.05662405979280534
+        }
+      },
+      "SPEC_all": {
+        "successes": 32,
+        "total": 32,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8928208017449293,
+          "upper": 1.0
+        }
+      },
+      "SPEC_valid": {
+        "successes": 32,
+        "total": 32,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8928208017449293,
+          "upper": 1.0
+        }
+      }
+    },
+    "QLORA": {
+      "AR_all": {
+        "successes": 32,
+        "total": 32,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8928208017449293,
+          "upper": 1.0
+        }
+      },
+      "AR_valid": {
+        "successes": 32,
+        "total": 32,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8928208017449293,
+          "upper": 1.0
+        }
+      },
+      "BA_all": {
+        "successes": 64,
+        "total": 64,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.9433759402071946,
+          "upper": 1.0
+        }
+      },
+      "BA_valid": {
+        "successes": 64,
+        "total": 64,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.9433759402071946,
+          "upper": 1.0
+        }
+      },
+      "EFR_C_valid": {
+        "successes": 16,
+        "total": 16,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8063923194655637,
+          "upper": 1.0000000000000002
+        }
+      },
+      "EFR_D_valid": {
+        "successes": 16,
+        "total": 16,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8063923194655637,
+          "upper": 1.0000000000000002
+        }
+      },
+      "FAR_valid": {
+        "successes": 0,
+        "total": 16,
+        "value": 0.0,
+        "wilson_95": {
+          "lower": 0.0,
+          "upper": 0.1936076805344365
+        }
+      },
+      "FFR": {
+        "successes": 0,
+        "total": 64,
+        "value": 0.0,
+        "wilson_95": {
+          "lower": 0.0,
+          "upper": 0.05662405979280534
+        }
+      },
+      "SPEC_all": {
+        "successes": 32,
+        "total": 32,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8928208017449293,
+          "upper": 1.0
+        }
+      },
+      "SPEC_valid": {
+        "successes": 32,
+        "total": 32,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8928208017449293,
+          "upper": 1.0
+        }
+      }
+    },
+    "SYSTEM": {
+      "AR_all": {
+        "successes": 30,
+        "total": 32,
+        "value": 0.9375,
+        "wilson_95": {
+          "lower": 0.7985287685584542,
+          "upper": 0.982689432968359
+        }
+      },
+      "AR_valid": {
+        "successes": 30,
+        "total": 32,
+        "value": 0.9375,
+        "wilson_95": {
+          "lower": 0.7985287685584542,
+          "upper": 0.982689432968359
+        }
+      },
+      "BA_all": {
+        "successes": 62,
+        "total": 64,
+        "value": 0.96875,
+        "wilson_95": {
+          "lower": 0.8930270822904169,
+          "upper": 0.9913878616538281
+        }
+      },
+      "BA_valid": {
+        "successes": 62,
+        "total": 64,
+        "value": 0.96875,
+        "wilson_95": {
+          "lower": 0.8930270822904169,
+          "upper": 0.9913878616538281
+        }
+      },
+      "EFR_C_valid": {
+        "successes": 16,
+        "total": 16,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8063923194655637,
+          "upper": 1.0000000000000002
+        }
+      },
+      "EFR_D_valid": {
+        "successes": 16,
+        "total": 16,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8063923194655637,
+          "upper": 1.0000000000000002
+        }
+      },
+      "FAR_valid": {
+        "successes": 0,
+        "total": 16,
+        "value": 0.0,
+        "wilson_95": {
+          "lower": 0.0,
+          "upper": 0.1936076805344365
+        }
+      },
+      "FFR": {
+        "successes": 0,
+        "total": 64,
+        "value": 0.0,
+        "wilson_95": {
+          "lower": 0.0,
+          "upper": 0.05662405979280534
+        }
+      },
+      "SPEC_all": {
+        "successes": 32,
+        "total": 32,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8928208017449293,
+          "upper": 1.0
+        }
+      },
+      "SPEC_valid": {
+        "successes": 32,
+        "total": 32,
+        "value": 1.0,
+        "wilson_95": {
+          "lower": 0.8928208017449293,
+          "upper": 1.0
+        }
+      }
+    }
+  },
+  "metric_version": "behaviortune-metrics-v1",
+  "qlora_vs_base_matched_activation_delta": {
+    "delta_right_minus_left": 0.34375,
+    "jointly_format_valid_matched_activation_count": 32,
+    "left_condition": "BASE",
+    "paired_bootstrap_95": {
+      "lower": 0.1875,
+      "upper": 0.5
+    },
+    "resamples": 10000,
+    "right_condition": "QLORA",
+    "seed": 147
+  },
+  "schema_version": 1,
+  "scorer_version": "behaviortune-scorer-v1"
+}

gate_decision.json

@@ -0,0 +1,54 @@
+{
+  "gate_scope": [
+    "delta_qlora_valid_minimum",
+    "delta_qlora_valid_bootstrap_lower_bound_gt",
+    "far_valid_qlora_maximum",
+    "far_valid_qlora_minus_base_maximum",
+    "ffr_qlora_maximum",
+    "ffr_base_maximum"
+  ],
+  "gates": {
+    "delta_qlora_valid_bootstrap_lower_bound_gt": {
+      "comparator": ">",
+      "observed": 0.1875,
+      "passed": true,
+      "threshold": 0.0
+    },
+    "delta_qlora_valid_minimum": {
+      "comparator": ">=",
+      "observed": 0.34375,
+      "passed": true,
+      "threshold": 0.2
+    },
+    "far_valid_qlora_maximum": {
+      "comparator": "<=",
+      "observed": 0.0,
+      "passed": true,
+      "threshold": 0.2
+    },
+    "far_valid_qlora_minus_base_maximum": {
+      "comparator": "<=",
+      "observed": 0.0,
+      "passed": true,
+      "threshold": 0.15
+    },
+    "ffr_base_maximum": {
+      "comparator": "<=",
+      "observed": 0.0,
+      "passed": true,
+      "threshold": 0.05
+    },
+    "ffr_qlora_maximum": {
+      "comparator": "<=",
+      "observed": 0.0,
+      "passed": true,
+      "threshold": 0.05
+    }
+  },
+  "result": "PASS",
+  "retry_executed": false,
+  "retry_trigger_observed": false,
+  "schema_version": 1,
+  "second_scientific_run_executed": false,
+  "stop_after_decision": true
+}

training_manifest.json

@@ -0,0 +1,77 @@
+{
+  "adapter_file_sha256": {
+    "adapter/README.md": "e38830bd0b1e50cd77eaa5101b7c378da55c6179b89cd3aa1f735dceea723744",
+    "adapter/adapter_config.json": "d6ebc4b0b2ada4e0020aef4f83d95f7eb4745defe9c4ecacf6d1f3aaca11111f",
+    "adapter/adapter_model.safetensors": "8d16ef2cb6ff7a982511fd58f21eff52538761f4d198b4cc5cbfd73ca7c9d4de",
+    "adapter/added_tokens.json": "c0284b582e14987fbd3d5a2cb2bd139084371ed9acbae488829a1c900833c680",
+    "adapter/merges.txt": "8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5",
+    "adapter/special_tokens_map.json": "76862e765266b85aa9459767e33cbaf13970f327a0e88d1c65846c2ddd3a1ecd",
+    "adapter/tokenizer.json": "aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4",
+    "adapter/tokenizer_config.json": "fed1bd317e1d5c4b32af689a7dec579e51392ba41942ec94a8dfba8ba85deefd",
+    "adapter/training_args.bin": "ff8df60789de5ccd595bef1bc046475c1d5407ce2df88ea24eaea862c0200064",
+    "adapter/vocab.json": "ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910"
+  },
+  "data_seed": 147,
+  "dependency_freeze_sha256": "c15f0bd5445c0b57670f8acfed0500ec6b970b5b20e6c8fe2b26cc3cfbf78486",
+  "dependency_versions": {
+    "accelerate": "1.6.0",
+    "bitsandbytes": "0.45.5",
+    "datasets": "3.6.0",
+    "huggingface_hub": "0.30.2",
+    "peft": "0.15.2",
+    "safetensors": "0.5.3",
+    "torch": "2.7.1+cu128",
+    "transformers": "4.51.3",
+    "trl": "0.16.1"
+  },
+  "epochs_completed": 3,
+  "gpu": "NVIDIA A100-SXM4-40GB",
+  "lora": {
+    "bias": "none",
+    "lora_alpha": 64,
+    "lora_dropout": 0.05,
+    "r": 32,
+    "target_modules": "all-linear",
+    "task_type": "CAUSAL_LM"
+  },
+  "optimizer_steps": 90,
+  "r1_train_sha256": "be16a7c198cacd727992f4ff82493ec40e5c042f9a485e2091a62eab213f91eb",
+  "recipe_id": "behaviortune-v1-qlora-primary-r1",
+  "recipe_sha256": "7e3852e87a59e9a5937e395e6cd2442142c99d2f6cac52576a1662b9e81251eb",
+  "retry_recipe_used": false,
+  "schema_version": 1,
+  "seed": 147,
+  "status": "PASS",
+  "train_metrics": {
+    "total_flos": 1998834477760512.0,
+    "train_loss": 0.004772673023767501,
+    "train_runtime": 195.2685,
+    "train_samples_per_second": 3.687,
+    "train_steps_per_second": 0.461
+  },
+  "training": {
+    "best_checkpoint_selection": false,
+    "bf16": true,
+    "data_seed": 147,
+    "early_stopping": false,
+    "effective_batch_size": 8,
+    "epochs": 3,
+    "final_checkpoint": "epoch_3",
+    "fp16": false,
+    "gradient_accumulation_steps": 4,
+    "gradient_checkpointing": true,
+    "learning_rate": 0.0002,
+    "lr_scheduler": "linear",
+    "max_gradient_norm": 1.0,
+    "max_sequence_length": 4096,
+    "optimizer": "adamw_torch",
+    "packing": false,
+    "per_device_train_batch_size": 2,
+    "seed": 147,
+    "shuffle_train_split": true,
+    "warmup_ratio": 0.05,
+    "weight_decay": 0.0
+  },
+  "training_rows": 240,
+  "training_run_count": 1
+}