Update README.md

README.md

@@ -31,7 +31,6 @@ tags:
 - calibration
 - system-one
 - multimodal
-- decision-model
 ---
 
 <div align="center">
@@ -50,22 +49,21 @@ tags:
 # d1-3B
 
 d1-3B is a 3B parameter **decision model** built on [LFM2.5-VL-3B](https://huggingface.co/LiquidAI/LFM2.5-VL-3B).
-You give it a state (text, JSON, images, or a mix) and a set of questions. It returns calibrated,
-typed answers in **one forward pass with zero output tokens**.
+You give it a state (text, JSON, images, or a mix) and a set of named questions. It returns calibrated,
+typed answers in **one forward pass with zero output tokens**: every answer is read directly from the
+model's distribution over the options, with no generation and no parsing.
 
-- **Best decision model under 10B on the Decision Index 0.2.1**: 48.57, ahead of every 4B and 9B model
-  and of Decider 35B-A3B (47.11).
-- **Multimodal**: images and text in the same state. It scores 74.1 on 11 public image benchmarks
+- **Best decision model under 10B**: 47.12 on the Decision Index 0.2.1. That is ahead of every 4B and 9B
+  model on the public leaderboard and level with Decider 35B-A3B (47.11).
+- **Multimodal**: images and text in the same state. It scores 73.7 on 11 public image benchmarks
   (LFM2.5-VL-3B: 73.9).
 - **Fast**: 8 ms a decision on an NVIDIA RTX 4090, 9 ms on an AMD MI325X, 30 ms on an Apple M5 Pro.
 
-Find more information about open d1 in our [blog post](https://www.liquid.ai/blog/open-d1).
-
-![image](https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/r2H7UlZ_m48SYAhZWIHiu.png)
+![Decision Index against model size](assets/di_pareto.png)
 
 > [!NOTE]
 > 💻 **Demos**: Try d1-3B in a Hugging Face space without any setup:
-> [**Open d1 Arcade**](https://huggingface.co/spaces/LiquidAI/system-one-arcade): Collection of 10 demos using d1-3B
+> **TODO: add demo**
 
 
 ## 🗒️ Model Details
@@ -73,14 +71,16 @@ Find more information about open d1 in our [blog post](https://www.liquid.ai/blo
 | Model | Parameters | Description |
 |---|---|---|
 | [LFM2.5-VL-3B](https://huggingface.co/LiquidAI/LFM2.5-VL-3B) | 3.1B | General-purpose vision-language model (base) |
-| **[d1-3B](https://huggingface.co/LiquidAI/d1-3B)** | 3.1B | Post-trained for single-pass, calibrated decisions |
+| **[d1-3B](https://huggingface.co/LiquidAI/d1-3b-RC)** | 3.1B | Post-trained for single-pass, calibrated decisions |
 
 d1-3B is a multimodal decision model with the following features:
 
 - **Total parameters**: 3.12B
+- **LM backbone**: LFM2.5-2.6B, 30 layers (22 double-gated short convolution blocks + 8 GQA)
 - **Vision encoder**: SigLIP2 NaFlex shape-optimized 400M
 - **Context length**: 32,768 tokens
 - **Vocabulary size**: 128,000
+- **Output**: no generated tokens. Each answer is a probability distribution over the question's options.
 
 We recommend d1-3B wherever a pipeline needs a yes/no, a pick from named options, or a rating:
 routing and triage, moderation, intent and topic classification, extraction checks, reranking, LLM-as-a-judge
@@ -97,47 +97,36 @@ pip install "transformers>=5.14" torch torchvision pillow
 The model ships its own code, so load it with `trust_remote_code=True`:
 
 ```python
+import io
+import urllib.request
+
 import torch
+from PIL import Image
 from transformers import AutoModel
-from transformers.image_utils import load_image
 
 device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
-dtype = torch.float32 if device == "cpu" else torch.bfloat16
-model = AutoModel.from_pretrained("LiquidAI/d1-3B", trust_remote_code=True, dtype=dtype).to(device)
+model = AutoModel.from_pretrained("LiquidAI/d1-3b-RC", trust_remote_code=True,
+                                  dtype=torch.float32 if device == "cpu" else torch.bfloat16).to(device)
 
-# Text: several named questions over one state, answered in one pass
+# Several named questions over one text state, answered in one pass
 questions = {
-    "refund": {
-        "type": "noul",
-        "instructions": "Is the customer asking for a refund?",
-    },
-    "team": {
-        "type": "choice",
-        "instructions": "Which team should handle this?",
-        "criteria": {
-            "billing": "Charges, refunds, invoices",
-            "technical": "App or site faults",
-            "fraud": "Suspected unauthorised use",
-        },
-    },
-    "urgency": {
-        "type": "score",
-        "instructions": "How urgent is this?",
-        "criteria": ["Can wait", "Today", "Blocking the customer now"],
-    },
+    "refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"},
+    "team": {"type": "choice", "instructions": "Which team should handle this?",
+             "criteria": {"billing": "Charges, refunds, invoices", "technical": "App or site faults",
+                          "fraud": "Suspected unauthorised use"}},
+    "urgency": {"type": "score", "instructions": "How urgent is this?",
+                "criteria": ["Can wait", "Today", "Blocking the customer now"]},
 }
 print(model.system_one("I was charged twice this month, please refund one of them.", questions))
 
-# Image: the photo is the whole state
-image = load_image("http://images.cocodataset.org/val2017/000000039769.jpg")  # two cats on a sofa
-cats = {
-    "type": "choice",
-    "instructions": "How many cats are there?",
-    "criteria": {"one": "One", "two": "Two", "more": "Three or more"},
-}
-print(model.system_one(None, {"cats": cats}, images=[image]))
+# An image as the whole state
+url = "http://images.cocodataset.org/val2017/000000039769.jpg"  # two cats on a sofa
+photo = Image.open(io.BytesIO(urllib.request.urlopen(url).read()))
+print(model.system_one(None, {"cats": {"type": "choice", "instructions": "How many cats are there?",
+                                       "criteria": {"one": "One", "two": "Two", "more": "Three or more"}}},
+                       images=[photo]))
 
-# Batch: many requests, packed together with no padding
+# Many requests, packed together with no padding
 tickets = ["Where is my parcel? It was due Monday.", "The app crashes when I open settings."]
 print(model.system_one_batch([(t, {"team": questions["team"]}) for t in tickets]))
 ```
@@ -163,32 +152,16 @@ Each call returns `{"answers": {name: answer}, "usage": {"input_tokens": n, "out
 
 ## ⚡ Speed
 
-Warm calls, one request at a time: a single question, three questions over one state, a 3.4k-token
-state and a 384 px image. The last column is throughput with 64 states packed into one pass.
-
-### Edge Inference
-
-We measure latency on an Apple M5 Pro and, in collaboration with NVIDIA, on an NVIDIA Jetson AGX Thor,
-a Jetson AGX Orin 64 GB and a Jetson Orin Nano.
+Warm calls in bf16, median of 20 runs:
 
 | | one question | 3 questions, one pass | 3.4k-token state | 384 px image | 64 states, packed |
 |---|---:|---:|---:|---:|---:|
-| Apple M5 Pro (`mps`) | 30 ms | 41 ms | 640 ms | 62 ms | 78 / s |
-| NVIDIA Jetson AGX Thor | 16 ms | 20 ms | 220 ms | 35 ms | 262 / s |
-| NVIDIA Jetson AGX Orin 64 GB | 26 ms | 35 ms | 560 ms | 83 ms | 110 / s |
-| NVIDIA Jetson Orin Nano | 50 ms | 73 ms | 1640 ms | 202 ms | 38 / s |
+| NVIDIA RTX 4090 | **8.0 ms** | 21 ms | 102 ms | 17 ms | 475 / s |
+| AMD MI325X | 9.2 ms | 14 ms | 44 ms | 18 ms | 1,106 / s |
+| Apple M5 Pro (`mps`) | 30 ms | 41 ms | 0.64 s | 62 ms | 78 / s |
 
-### GPU Inference
-
-We measure latency on an NVIDIA RTX 4090 and an AMD MI325X, in bf16, median of 20 runs.
-
-| | one question | 3 questions, one pass | 3.4k-token state | 384 px image | 64 states, packed |
-|---|---:|---:|---:|---:|---:|
-| NVIDIA RTX 4090 | 8 ms | 21 ms | 102 ms | 17 ms | 475 / s |
-| AMD MI325X | 9 ms | 14 ms | 44 ms | 18 ms | 1,106 / s |
-
-On NVIDIA GPUs, `model.compile(mode="reduce-overhead")` runs single questions as CUDA graphs (the RTX 4090
-row uses it). Without it, a single question takes 16 ms. The first call with a new shape pays for kernel
+On NVIDIA GPUs, `model.compile(mode="reduce-overhead")` runs single questions as CUDA graphs. The RTX 4090
+row uses it; without it a single question takes 16 ms. The first call with a new shape pays for kernel
 selection or compilation, so warm up the shapes you serve.
 
 ## 📊 Performance
@@ -197,12 +170,15 @@ All results are on public benchmarks.
 
 ### Decision Index 0.2.1
 
-We scored d1-3B with the official scorer (not a leaderboard submission). All other rows come from the public leaderboard v0.2.1.
+The [Decision Index](https://huggingface.co/spaces/multimodalart/jev-decision-index) casts 38 public
+benchmarks as 120,226 typed decisions, scored as chance-corrected skill. We scored d1-3B with the official
+scorer; it is not a leaderboard submission. All other rows come from the public leaderboard
+(snapshot 2026-09-28). Sizes are served parameters.
 
 | Model | Size | Decision Index | Knowledge | Language | Retrieval | Tools | Arts |
 |---|---:|---:|---:|---:|---:|---:|---:|
 | Winnow-12B | 12B | 50.02 | 33.8 | 56.0 | 54.0 | 71.0 | 30.0 |
-| **d1-3B** | **3B** | **48.57** | 23.8 | 56.4 | 52.8 | **74.5** | **36.3** |
+| **d1-3B** | **3B** | **47.12** | 27.8 | 53.5 | 52.2 | 66.7 | **34.6** |
 | Decider 35B-A3B | 36B | 47.11 | 31.8 | 55.5 | 54.7 | 56.5 | 32.6 |
 | JPT-9B | 9.7B | 46.89 | 31.7 | 56.7 | 44.6 | 67.0 | 28.6 |
 | Decision 1.0 Lux | 9.7B | 43.49 | 30.9 | 48.0 | 50.0 | 57.2 | 26.4 |
@@ -214,22 +190,23 @@ We scored d1-3B with the official scorer (not a leaderboard submission). All oth
 
 ### Benchmarks as decisions
 
-Besides the Decision Index, we added a few other internal evaluations based on public benchmarks.
+Evaluation splits are cast as typed decisions, with the same rows for every model. These scores follow the
+decision protocol, not each benchmark's own.
 
 | Benchmark | d1-3B | Decider 4B | Decider 2B |
 |---|---:|---:|---:|
-| SQuAD 2.0 | **85.3** | 76.0 | 67.7 |
-| Civil Comments | 93.0 | 92.8 | **93.6** |
-| MASSIVE intent | 87.3 | **88.3** | 81.1 |
-| HelpSteer2 | 36.7 | **42.0** | 32.0 |
-| PubMedQA | **66.0** | 63.3 | 65.7 |
-| BoolQ | 86.7 | **89.0** | 87.3 |
-| XNLI | 85.0 | **88.6** | 85.0 |
-| PAWS-X | **76.9** | 69.8 | 59.5 |
-| **Mean** | **77.1** | 76.2 | 71.5 |
-
-d1-3B also scores 71.8 on [DecisionBench](https://huggingface.co/datasets/Hanno-Labs/decision-bench) (eng v1,
-all 23,900 rows) and 69.3 on [Fast Decisions](https://huggingface.co/datasets/fastino/fast-decisions)
+| SQuAD 2.0 | **83.3** | 76.0 | 67.7 |
+| Civil Comments | 93.3 | 92.8 | **93.6** |
+| MASSIVE intent | 86.9 | **88.3** | 81.1 |
+| HelpSteer2 | 38.0 | **42.0** | 32.0 |
+| PubMedQA | **68.3** | 63.3 | 65.7 |
+| BoolQ | 86.3 | **89.0** | 87.3 |
+| XNLI | 85.6 | **88.6** | 85.0 |
+| PAWS-X | **76.4** | 69.8 | 59.5 |
+| **Mean** | **77.3** | 76.2 | 71.5 |
+
+d1-3B also scores 69.5 on [DecisionBench](https://huggingface.co/datasets/Hanno-Labs/decision-bench) (eng v1,
+all 23,900 rows) and 68.8 on [Fast Decisions](https://huggingface.co/datasets/fastino/fast-decisions)
 (dev split).
 
 ### Vision
@@ -239,22 +216,29 @@ compared with the base model:
 
 | Benchmark | d1-3B | LFM2.5-VL-3B |
 |---|---:|---:|
-| AI2D | 79.9 | 80.9 |
-| BLINK | 59.2 | 58.7 |
-| CV-Bench | 82.1 | 87.6 |
-| HallusionBench | 65.3 | 65.0 |
-| MMBench | 84.9 | 84.3 |
-| MME | 82.1 | 82.4 |
-| MMStar | 59.9 | 61.2 |
-| MMVP | 77.0 | 73.7 |
+| AI2D | 80.1 | 80.9 |
+| BLINK | 58.2 | 58.7 |
+| CV-Bench | 82.7 | 87.6 |
+| HallusionBench | 63.8 | 65.0 |
+| MMBench | 84.3 | 84.3 |
+| MME | 81.7 | 82.4 |
+| MMStar | 59.4 | 61.2 |
+| MMVP | 76.3 | 73.7 |
 | POPE | 88.5 | 90.1 |
-| VisualWebBench | 71.4 | 78.3 |
-| VL-RewardBench | 65.0 | 50.9 |
-| **Mean** | **74.1** | 73.9 |
-| [ImajevBench](https://huggingface.co/datasets/mohit67890/imajev-bench) (dev and calibration, 253 rows) | 64.0 | 66.8 |
+| VisualWebBench | 71.6 | 78.3 |
+| VL-RewardBench | 63.6 | 50.9 |
+| **Mean** | **73.7** | 73.9 |
+| [ImajevBench](https://huggingface.co/datasets/mohit67890/imajev-bench) (dev and calibration, 253 rows) | 66.8 | 66.8 |
 
 With the images removed, the same questions score 45.1, so the answers come from the images.
 
+## ⚠️ Limitations
+
+- **Knowledge-heavy questions**: larger models stay ahead on hard knowledge (for example MMLU-Pro and GPQA).
+- **Thresholds**: calibrate any threshold on data from your own deployment before gating on it.
+- **No text output**: d1-3B answers questions over fixed options. Use [LFM2.5-VL-3B](https://huggingface.co/LiquidAI/LFM2.5-VL-3B)
+  for open-ended generation.
+
 ## 📬 Contact
 
 - Got questions or want to connect? [Join our Discord community](https://discord.com/invite/liquid-ai)
@@ -262,16 +246,6 @@ With the images removed, the same questions score 45.1, so the answers come from
 
 ## Citation
 
-```bibtex
-@article{liquidAI2026opend1,
-  author  = {Liquid AI},
-  title   = {Open d1: Edge decision models for text, vision, and audio},
-  journal = {Liquid AI Blog},
-  year    = {2026},
-  note    = {https://www.liquid.ai/blog/d1-open},
-}
-```
-
 ```bibtex
 @article{liquidai2025lfm2,
   title   = {LFM2 Technical Report},

assets/d1-3b-smash.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08f69b1a67bf32cd18fad32b5872519b580be9e94bafbe26a13d5cf0aaa49c1f
+size 53062187

assets/ood_tasks.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b491dfa851157e12ba2bee33a394c9565408247cf0004e493730622c62878e42
+size 8931625

config.json

@@ -3,7 +3,7 @@
     "Lfm2VlForConditionalGeneration"
   ],
   "auto_map": {
-    "AutoModel": "modeling_d1.D1Model"
+    "AutoModel": "modeling_lfm_jev.LfmJevModel"
   },
   "bos_token_id": 124894,
   "do_image_splitting": true,

lfm2_vl.py

@@ -1,9 +1,9 @@
 """LFM2-VL and LFM2 on the hybrid stack.
 
 HF's classes stay the frame (the SigLIP2 tower, the projector, the image
-features, loading and saving); only the language model is replaced, with HF's
-module names, so checkpoints load and save unchanged. A state with any number
-of questions runs as one `hybrid.Tree`, with the same outputs as HF's own
+features, loading and saving); only the language model is ours, with HF's module
+names, so checkpoints load and save unchanged, and a state
+with any number of questions runs as one `hybrid.Tree`, which matches HF's own
 implementation.
 """
 
@@ -46,7 +46,7 @@ class Attention(nn.Module):
 
 
 class ShortConv(nn.Module):
-    """`out(C * conv(B * x))`; HF's `nn.Conv1d` holds the taps, `causal_conv` runs them."""
+    """`out(C * conv(B * x))`; HF's `nn.Conv1d` holds the taps, our causal conv runs them."""
 
     def __init__(self, cfg):
         super().__init__()

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:50e03317847caf6df9a9aee27ed40f20554a86a21e60d1d47ba41a422b546c0c
+oid sha256:042be2b86f2140cf7e696d9d26c0dd8d2f0f8f943b783eb9ffa34903b41c98bb
 size 6247065504

modeling_d1.py → modeling_lfm_jev.py

@@ -14,7 +14,7 @@ from .lfm2_vl import Lfm2VlForConditionalGeneration
 from .runner import SystemOne
 
 
-class D1Model(Lfm2VlForConditionalGeneration):
+class LfmJevModel(Lfm2VlForConditionalGeneration):
     @cached_property
     def engine(self) -> SystemOne:
         from transformers import AutoTokenizer

prompt.py

@@ -20,6 +20,7 @@ SYSTEMS: dict[str, str | None] = {
     ),
 }
 DEFAULT_SYSTEM = "none"
+SYSTEM = SYSTEMS[DEFAULT_SYSTEM]
 
 IM_START = "<|im_start|>"
 IM_END = "<|im_end|>"
@@ -54,6 +55,20 @@ class Score:
 Question = Choice | Noul | Score
 
 
+def row_to_question(row: dict) -> Question:
+    """A row's question (`kind`, `instructions`, `criteria`), as the model is asked it."""
+    kind = row["kind"]
+    if kind == "noul":
+        return Noul(instructions=row["instructions"], criteria=row.get("criteria"))
+    if kind == "score":
+        return Score(instructions=row["instructions"], criteria=list(row["criteria"]))
+    return Choice(instructions=row["instructions"], criteria=row["criteria"])
+
+
+def cardinality(q: Question) -> int:
+    return 2 if isinstance(q, Noul) else len(q.criteria)
+
+
 # --------------------------------------------------------------------------- #
 # verbalizer
 # --------------------------------------------------------------------------- #
@@ -109,9 +124,12 @@ _TOKENIZERS: dict[int, object] = {}
 def aliases(tokenizer, labels: Sequence[str]) -> list[tuple[str, int]]:
     """Assign every label a distinct single-token code: [(code, token_id)].
 
-    Memoised on the codes rather than on the labels: the codes are positional
-    unless the labels are already letters, so every option list of one length
-    shares an entry.
+    Memoised on the codes rather than on the labels. The codes are positional
+    unless the labels are already letters, so two orderings of one option list
+    share an entry, and a 151-option inventory question does not re-derive its
+    codes once per permuted row. Keying on the label tuple instead misses on
+    every such row. Measured against a contended node it was not the cost it
+    first looked like, but the entry is the codes and the key should say so.
     """
     _TOKENIZERS.setdefault(id(tokenizer), tokenizer)
     codes = tuple(option_codes(labels))
@@ -191,16 +209,36 @@ def readout(tokenizer, q: Question, logz, calibration=None) -> list[float]:
     return [e / sum(exps) for e in exps]
 
 
+def option_labels(q: Question) -> list[str]:
+    if isinstance(q, Noul):
+        return ["yes", "no"]
+    if isinstance(q, Score):
+        return [str(i) for i in range(len(q.criteria))]
+    return list(q.criteria.keys())
+
+
+def gold_code(tokenizer, q: Question, gold: str | int) -> str:
+    """The exact string the model must produce in the answer slot."""
+    if isinstance(q, Noul):
+        if isinstance(gold, str):
+            return "yes" if gold.strip().lower() in {"yes", "true", "1"} else "no"
+        return "yes" if int(gold) == 1 else "no"
+    if isinstance(q, Score):
+        return str(int(gold))
+    labels = list(q.criteria.keys())
+    idx = labels.index(gold) if gold in labels else int(gold)
+    return aliases(tokenizer, labels)[idx][0]
+
+
 # --------------------------------------------------------------------------- #
 # state and question rendering
 # --------------------------------------------------------------------------- #
 
 
-DEFAULT_MODEL = "LiquidAI/LFM2.5-VL-3B"
-
 # How a state is rendered: `json_only`, the default, writes every state as the object it is; `json` keeps
 # three shortcuts (`Message:`, `Passage:` / `Asked:`, a lone question's text); `sections` writes nested
 # states as labelled blocks.
+DEFAULT_MODEL = "LiquidAI/LFM2.5-VL-3B"
 DEFAULT_STATE_STYLE = "json_only"
 
 
@@ -211,8 +249,9 @@ def _is_scalar(v: Any) -> bool:
 def _sections(obj: Any, path: str, out: list[str]) -> None:  # noqa: C901
     """Flatten a nested state into labelled blocks, keeping real newlines.
 
-    `json.dumps` escapes every newline inside a log line or a record, so a
-    multi-line record would arrive as one string of `\n`; this keeps it readable.
+    A caller's state is JSON, but `json.dumps` escapes every newline inside a
+    log line or a record, so a 40-line CrowdStrike host record arrives as one
+    unreadable string of `\n`. The model reads a case file, not a payload.
     """
     head = f"[{path}]\n" if path else ""
     if isinstance(obj, dict):
@@ -319,8 +358,8 @@ def prefix_text(
     return f"{bos}{turn}{IM_START}user\n{images}{body}"
 
 
-# What sits between the assistant header and the answer slot, per model type: nothing on LFM2-VL, whose
-# template opens no reasoning block.
+# What sits between the assistant header and the answer slot: nothing on LFM2-VL, whose template opens
+# no reasoning block, so the slot is already clean. A family module can add its own (`LEADS`).
 DEFAULT_LEAD = ""
 LEADS: dict[str, str] = {}
 

runner.py

@@ -39,7 +39,7 @@ MODELS = {
 
 
 def load_backbone(model_id: str = DEFAULT_MODEL, dtype=torch.bfloat16):
-    """The checkpoint and its tokenizer, with SDPA attention."""
+    """The checkpoint and its tokenizer. SigLIP2 has no FA2 kernel here, so SDPA."""
     from transformers import AutoModelForImageTextToText, AutoTokenizer, PretrainedConfig
 
     tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
@@ -82,8 +82,8 @@ class SystemOne(SystemOneApi):
         model=None,
         tokenizer=None,
     ):
-        """`model` and `tokenizer`, when given, are a backbone already loaded (`D1Model` passes itself);
-        it stays on its device unless `device` says otherwise."""
+        """`model` and `tokenizer`, when given, are a backbone already loaded (the Hub's remote-code class
+        passes itself); it stays on its device unless `device` says otherwise."""
         if model is None:
             model, tokenizer = load_backbone(model_id)
         else:
@@ -103,7 +103,8 @@ class SystemOne(SystemOneApi):
         self.option_style = option_style
         self.token_budget = token_budget
         self.processor = None
-        # CUDA graphs for single questions on NVIDIA, where eager time is mostly kernel launches.
+        # CUDA graphs for single questions, on NVIDIA: where eager time is kernel launches, 16 to 8 ms a
+        # decision on an RTX 4090. The first two prompt lengths compile; later ones reuse the dynamic graph.
         if compile and torch.version.hip:
             raise ValueError("compile=True needs CUDA: on ROCm the CUDA graphs fault after a few dozen calls")
         self._one_pass = (torch.compile(self.model.forward, mode="reduce-overhead")
@@ -119,14 +120,19 @@ class SystemOne(SystemOneApi):
 
     # --------------------------------------------------------------- forward
 
-    def _logz_ids(self, rows: list[list[int]]) -> list[torch.Tensor]:
-        """Log-softmax at the answer slot, one row per token list, in one pass.
+    @torch.inference_mode()
+    def _logz(self, texts: Sequence[str], questions: Sequence[Question] | None = None) -> list[torch.Tensor]:
+        """Log-softmax at the answer slot, one row per text, in one pass; the slot needs only the texts, and
+        `questions` is accepted for the scoring interface.
 
         The rows are one tree (`hybrid.py`): their common start is its trunk and
         is read once; the rest of each row is a branch, packed with no padding.
         Mathematically each row alone; in bf16 the kernels differ by batch shape.
         """
-        if len(rows) == 1:  # nothing to share: a plain chain is faster than a tree of one
+        return self._logz_ids([self.tokenizer.encode(t, add_special_tokens=False) for t in texts])
+
+    def _logz_ids(self, rows: list[list[int]]) -> list[torch.Tensor]:
+        if len(rows) == 1:  # nothing to share: a plain chain, 6-12% faster than a tree of one
             row = self._one_pass(input_ids=torch.tensor(rows, device=self.device), logits_to_keep=1).logits[0, -1]
             return [row.float() - torch.logsumexp(row.float(), dim=-1)]
         shared = 0  # every row keeps at least its last token