Add jeba-en LoRA adapter, calibration temperature, and model card

README.md

@@ -0,0 +1,85 @@
+---
+license: apache-2.0
+library_name: jeba
+language:
+  - en
+  - pt
+  - es
+  - fr
+  - de
+  - multilingual
+tags:
+  - decision-engine
+  - system-one
+  - calibration
+  - multilingual
+  - local-first
+pipeline_tag: text-classification
+---
+
+# jeba-en (System One decision engine)
+
+> **Status: pre-release.** This card describes the intended release. **Weights and measured
+> metrics are pending** the RTX 3060 training run (`uv sync --extra train`); numbers below are
+> placeholders to be replaced by `benchmarks/report.md`.
+
+## Model details
+
+- **Developed by:** The jeba Authors.
+- **Model type:** non-autoregressive encoder with three task distributions (`noul`, `choice`,
+  `score`), answering typed questions about a state in one forward pass.
+- **Trunk:** ModernBERT-large (English) and mmBERT-base (100+ languages); see ADR-0007.
+- **Licence:** Apache-2.0.
+- **Repository:** <https://github.com/munod/jeba>
+
+## Uses
+
+jeba answers atomic `choice` / `score` / `noul` questions about a state and returns typed values
+with probabilities and `confidence`. It speaks the TypeSafe Jev `/v1/systemone` wire protocol as
+a drop-in and runs **locally/offline** with no API key. Compose several atomic answers in code
+rather than asking one broad question.
+
+**Out of scope:** free-form text generation, multi-step reasoning, and any decision requiring
+extended deliberation — decompose those into atomic questions and combine results in code.
+
+## Bias, risks, and limitations
+
+- Probabilities are only meaningful after **calibration**; the shipped temperature must be
+  applied (see `docs/training.md`).
+- Synthetic training data can inherit generator biases; public probes are evaluation-only.
+- Confidence is a property of the distribution, not a guarantee of correctness.
+
+## Training
+
+Deterministic synthetic JSONL (`training/generate_data.py`) supervised with an RLCD
+proper-scoring objective (`training/finetune_rlcd.py`), then temperature-calibrated on a held-out
+split (`training/fit_calibration.py`). Configs and seed live under `training/configs/`.
+
+## Evaluation
+
+Reported by `training/evaluate.py` and rendered by `benchmarks/report.py` (accuracy, ECE, p50/p95
+latency per primitive and language).
+
+**Full-scale run (single RTX 3060 12GB):** 6,000 train / 1,500 eval deterministic synthetic
+records, LoRA (r=16), 3 epochs, batch 16, bf16 + gradient checkpointing.
+
+| Checkpoint | Accuracy | ECE (calibrated) | p50 (ms) |
+| --- | --- | --- | --- |
+| English (ModernBERT-large + LoRA) | 0.536 | 0.090 | 17.6 |
+| Multilingual (mmBERT-base + LoRA) | 0.533 | 0.016 | 12.2 |
+
+Per primitive (English): `noul` 0.726 acc, `score` 0.574, `choice` 0.308. Labels are synthetic
+and template-limited (choice ~0.25, score ~0.25, noul ~0.5 chance), so `choice` is near chance;
+richer data should lift it. Full tables and environment are in
+[`benchmarks/report.md`](https://github.com/munod/jeba/blob/main/benchmarks/report.md).
+
+## Citation
+
+```bibtex
+@misc{jeba2026,
+  title        = {jeba: a local-first System One decision engine},
+  author       = {The jeba Authors},
+  year         = {2026},
+  howpublished = {\url{https://github.com/munod/jeba}}
+}
+```

adapter_config.json

@@ -0,0 +1,50 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": {
+    "base_model_class": "ModernBertModel",
+    "parent_library": "transformers.models.modernbert.modeling_modernbert"
+  },
+  "base_model_name_or_path": "answerdotai/ModernBERT-large",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "kasa_config": null,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "monteclora_config": null,
+  "peft_type": "LORA",
+  "peft_version": "0.21.0",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "Wqkv",
+    "Wi",
+    "Wo"
+  ],
+  "target_parameters": null,
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false,
+  "velora_config": null
+}

adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9eca12a7d798d7c4442ace62b4051f8ea09a61eaddbe0caed8b098206685dbc
+size 28813352

finetune_config.json

@@ -0,0 +1,19 @@
+{
+  "batch_size": 16,
+  "bf16": true,
+  "data_path": "data/train_en.jsonl",
+  "epochs": 3,
+  "grad_accum": 4,
+  "gradient_checkpointing": true,
+  "learning_rate": 0.0001,
+  "lora_alpha": 32,
+  "lora_dropout": 0.05,
+  "lora_rank": 16,
+  "max_len": 256,
+  "max_records": null,
+  "model_id": "answerdotai/ModernBERT-large",
+  "out_dir": "checkpoints/en",
+  "seed": 42,
+  "use_4bit": false,
+  "val_split": 0.1
+}

temperature_calibration.json

@@ -0,0 +1,24 @@
+{
+  "bins": 10,
+  "per_primitive": {
+    "choice": {
+      "ece_after": 0.009169,
+      "ece_before": 0.034399,
+      "n": 500,
+      "temperature": 1.5
+    },
+    "noul": {
+      "ece_after": 0.09859,
+      "ece_before": 0.09859,
+      "n": 500,
+      "temperature": 1.0
+    },
+    "score": {
+      "ece_after": 0.190815,
+      "ece_before": 0.190815,
+      "n": 500,
+      "temperature": 1.0
+    }
+  },
+  "warnings": []
+}