README.md

---
license: apache-2.0
language:
- en
pipeline_tag: zero-shot-classification
tags:
- decision-model
- system-one
- falcondec
- arthur
- lightdec_arthur
- calibrated-decisions
- multiple-choice
- intent-classification
- customer-support
- natural-language-inference
- code
- guardrails
- agents
- selective-prediction
- byte-level
- falconsai
- model-surgeon
- attested-lineage
base_model: Falconsai/LightDec_Arthur
---

> Source model card: `Falconsai/LightDec_Arthur` @ `main`, carried verbatim below. Its licence is the repository's. The Model Surgeon record follows it.

# LightDec_Arthur

**[View in Model Surgeon](https://surgeon.falcons.ai/?hub=Falconsai/LightDec_Arthur)**

**LightDec_Arthur** is a small, fast, calibrated *decision model* from [Falcons.ai](https://huggingface.co/Falconsai). Given a piece of state (an email, a ticket, an agent trace, a log, a JSON record) and a typed question, it picks an answer from a closed set of options in a single forward pass and returns calibrated probabilities, so your application knows when to trust the answer and when to defer to a human or a larger model.

Unlike [Falconsai/LightDec_V2](https://huggingface.co/Falconsai/LightDec_V2), it has **no pretrained backbone and no tokenizer**. It reads raw UTF-8 bytes and was trained from scratch on the FalconDec_V2 decision mixture. At about 12 million parameters it is roughly 13× smaller than LightDec_V2, trading accuracy for size and speed.

| | |
| --- | --- |
| **Architecture** | Arthur (byte-level recursive encoder + option-interaction head), trained from scratch |
| **Backbone** | None |
| **Parameters** | 12,268,033 (fp16 weights; model folder 24.5 MB) |
| **Context** | 1,024 tokens of 4 bytes (4,096 bytes) per call; 2,048 tokens for questions with more than 24 options |
| **Question types** | `choice`, `noul` (yes/no), `score` (ordinal levels) |
| **Options per question** | 2 or more in one pass (up to 77 in the test set) |
| **Test accuracy** | 59.6% micro / 58.4% macro over 26,597 test items, 58 tasks |
| **Calibration** | ECE 0.030 after temperature scaling (0.246 before) |
| **Latency** | 5.1 ms per decision on GPU (NVIDIA RTX PRO 6000 Blackwell); CPU not yet measured |
| **Version** | Arthur notebook v0.8.0, base tier |

## What it does

LightDec_Arthur answers the same kind of small, structured questions as LightDec: *Which queue does this ticket go to? Is this code a clone of that one? Is this prompt a jailbreak? Which intent is this request?* It is not a generative model and never produces free text; it only chooses among the options you give it.

Three question types are supported:

- **`choice`**: pick one of N options.
- **`noul`**: a yes/no judgment (options `Yes`, `No`).
- **`score`**: an ordinal scale such as `["Calm", "Annoyed", "Furious"]`, trained with a ranked-probability term so near misses cost less than far ones.

The calibrated top probability works as a confidence. With a threshold of **0.80**, the model answered **21.5%** of test questions on its own and was **89.7%** accurate on those; **94.6%** of its mistakes fell in the deferred remainder.

## How it works

The question, the options and the state are packed into one byte sequence:

```
[CLS] question [SEP] [OPT] option_1 [OPT] option_2 ... [OPT] option_k [SEP] state [SEP]
```

Each byte gets an embedding, plus hashed embeddings of the 2-, 3- and 4-byte fragments ending at it (65,536 buckets). A depthwise convolution mixes neighbouring bytes, and every 4 bytes are pooled into one position; each `[OPT]` marker starts a new position. One transformer block (8 heads, rotary positions) is applied **6 times** over the text, with a learned step vector added before each pass.

The vector at each option's `[OPT]` position, plus the mean of the text and a question-type embedding, forms the option vector. The same block is applied **twice** more over the option vectors, without position encoding, so options can compare themselves with each other. A small MLP gives one logit per option. A temperature per question type and option-count bucket (2, 3–5, 6–12, 13+ options), fitted on validation data and stored in `config.json`, calibrates the softmax.

## Usage

This repository does not yet include a standalone loader. The model definition (`Arthur`), the byte layout (`assemble`) and the temperature step live in the Arthur training notebook. To score a decision with the notebook's code:

```python
import json, numpy as np, torch
from safetensors.torch import load_file

cfg = json.load(open("config.json"))                      # tier config, byte layout, temperatures
net = Arthur(TIER_CFG["base"])                             # class from the Arthur notebook
net.load_state_dict({k: v.float() for k, v in load_file("model.safetensors").items()})
net.to(DEVICE).eval()

d = dict(state="I was billed twice for my Pro plan in March. Please refund the duplicate charge.",
         question="What is this support email primarily about?",
         options=["billing", "technical", "feature request", "account"], qtype="choice")
ids, win, wmask, qt = batch_tensors([d])                   # from the Arthur notebook
with torch.no_grad():
    z = net(ids, win, wmask, qt)[0, :len(d["options"])].float().cpu().numpy()
t = np.array(cfg["temperatures"])[QTYPES[d["qtype"]], bucket(len(d["options"]))]
p = np.exp((z - z.max()) / t); p /= p.sum()
print(d["options"][int(p.argmax())], round(float(p.max()), 3), "DEFER" if p.max() < 0.8 else "")
```

## Evaluation

All numbers come from `results_full.json`, written by the training run. The test split has 26,597 items across 58 tasks in 9 domains. Tasks marked *held-out* have no training or validation data.


### By domain

Accuracy is the mean over the domain's tasks.

| Domain | Tasks | Test items | Accuracy |
| --- | --- | --- | --- |
| support | 2 | 1,000 | 99.5% |
| code | 9 | 3,854 | 74.1% |
| guardrails | 4 | 1,294 | 69.0% |
| agentic | 4 | 925 | 60.5% |
| workflows | 4 | 2,000 | 60.4% |
| intents | 4 | 1,800 | 58.3% |
| policy | 9 | 4,500 | 56.5% |
| reasoning | 18 | 9,224 | 47.2% |
| classification | 4 | 2,000 | 43.0% |

### Head-to-head against Arthur nano

Arthur nano is the 1.6 MB tier of the same architecture (793,729 parameters), scored on the same test set. It was trained for 8 epochs with a 512-token window, so this comparison mixes model size, epochs and input length.

On the 58 shared tasks, LightDec_Arthur averaged **58.4%** against **53.6%** for Arthur nano, winning on 39 tasks, tying on 3 and losing on 16.

<details>
<summary>Per-task head-to-head</summary>

| Task | n | Arthur | Arthur nano | Δ |
| --- | --- | --- | --- | --- |
| `policy/return_window_transfer` | 500 | 82.6% | 31.4% | +51.2 |
| `mbpp/solution` | 500 | 79.2% | 49.2% | +30.0 |
| `codexglue/doc_to_code` | 504 | 88.5% | 63.7% | +24.8 |
| `qasc/mcq` | 500 | 46.8% | 22.4% | +24.4 |
| `codexglue/code_to_doc` | 504 | 83.9% | 64.9% | +19.1 |
| `codexglue/func_name` | 467 | 69.8% | 51.8% | +18.0 |
| `hotpotqa/retrieve` | 497 | 58.6% | 41.6% | +16.9 |
| `hotpotqa/comparison_yes_no` | 26 | 65.4% | 50.0% | +15.4 |
| `sciq/mcq` | 498 | 60.8% | 47.6% | +13.3 |
| `banking77/intent` | 500 | 58.2% | 46.4% | +11.8 |
| `massive_en/intent` | 500 | 85.4% | 75.6% | +9.8 |
| `policy/refund_approval_transfer` | 500 | 52.0% | 42.6% | +9.4 |
| `snli/nli` | 988 | 72.4% | 63.8% | +8.6 |
| `snli/contradicts` | 500 | 85.4% | 77.4% | +8.0 |
| `policy/invoice_overdue_transfer` | 500 | 48.8% | 41.0% | +7.8 |
| `snli/must_be_true` | 500 | 90.8% | 83.6% | +7.2 |
| `policy/sla_urgency_transfer` | 500 | 30.2% | 23.0% | +7.2 |
| `openbookqa/mcq` | 500 | 36.2% | 30.2% | +6.0 |
| `policy/table_extreme_transfer` | 500 | 28.4% | 22.4% | +6.0 |
| `arc_challenge/mcq` | 500 | 29.4% | 23.8% | +5.6 |
| `winogrande/blank` | 500 | 56.2% | 50.8% | +5.4 |
| `sst5/score` | 500 | 27.4% | 22.6% | +4.8 |
| `clinc150/intent` | 500 | 84.8% | 80.2% | +4.6 |
| `policy/free_shipping_transfer` | 500 | 80.0% | 75.6% | +4.4 |
| `humaneval/completion` | 119 | 42.0% | 37.8% | +4.2 |
| `mnli/claim` | 500 | 52.0% | 48.6% | +3.4 |
| `typed_decisions/agent_trace_observability` | 500 | 69.8% | 66.6% | +3.2 |
| `commonsense_qa/mcq` | 493 | 19.1% | 16.8% | +2.2 |
| `boolq/yes_no` | 500 | 62.6% | 60.4% | +2.2 |
| `arc_easy/mcq` | 500 | 30.4% | 28.4% | +2.0 |
| `jailbreak/detect` | 262 | 91.2% | 89.3% | +1.9 |
| `devign/vulnerability` | 500 | 64.2% | 62.4% | +1.8 |
| `hellaswag/continuation` | 500 | 28.0% | 26.2% | +1.8 |
| `banking77/intent_77` | 300 | 4.7% | 3.0% | +1.7 |
| `civil_comments/toxic` | 500 | 86.8% | 85.6% | +1.2 |
| `ag_news/topic` | 500 | 84.6% | 83.6% | +1.0 |
| `bitext/route` | 500 | 99.4% | 98.8% | +0.6 |
| `bigclonebench/clone` | 500 | 95.4% | 95.0% | +0.4 |
| `mbpp/bugspot` | 256 | 43.8% | 43.4% | +0.4 |
| `codexglue/lang_id` | 504 | 99.8% | 99.8% | +0.0 |
| `policy/access_control_transfer` | 500 | 100.0% | 100.0% | +0.0 |
| `policy/invoice_total_transfer` | 500 | 52.0% | 52.0% | +0.0 |
| `scitail/support` | 500 | 66.0% | 66.2% | -0.2 |
| `bitext/category` | 500 | 99.6% | 99.8% | -0.2 |
| `mmlu/mcq` | 500 | 28.4% | 28.8% | -0.4 |
| `gsm8k/math` | 500 | 30.6% | 31.2% | -0.6 |
| `yelp/score` | 500 | 48.6% | 50.6% | -2.0 |
| `typed_decisions/security_incidents` | 500 | 65.8% | 68.2% | -2.4 |
| `typed_decisions/invoice_processing` | 500 | 56.0% | 59.4% | -3.4 |
| `policy/count_threshold_transfer` | 500 | 34.8% | 38.2% | -3.4 |
| `agentharm/refuse` | 416 | 50.7% | 54.3% | -3.6 |
| `aqua_rat/math` | 247 | 21.9% | 27.5% | -5.7 |
| `counsel/step_has_error` | 201 | 72.1% | 78.1% | -6.0 |
| `anli/nli` | 498 | 32.1% | 38.1% | -6.0 |
| `counsel/critique_quality` | 201 | 45.8% | 53.2% | -7.5 |
| `emotion/6way` | 500 | 11.6% | 19.2% | -7.6 |
| `typed_decisions/customer_service` | 500 | 50.0% | 58.0% | -8.0 |
| `prompt_injections/detect` | 116 | 47.4% | 56.0% | -8.6 |

</details>

### Input length

At the 1,024-token window, 385 of 972,625 training and test decisions (0.04%) have text cut off, concentrated in `counsel/critique_quality` (7.3%), `jailbreak/detect` (7.0%) and `counsel/step_has_error` (4.6%).

### Decision Index 0.2 (partial)

LightDec_Arthur was scored on 39 benchmarks of the public [Decision Index 0.2](https://github.com/webdevtodayjason/decision-index) with the index kit's own runner and scorers. Five benchmarks were not built (ToolRet and BRIGHT need `tiktoken`, POP909 needs `mido`, ChessBench needs `python-chess`, HLE is gated), so **this is not a Decision Index score**. The kit's rules apply: nothing is cut, and a request longer than 8,192 bytes, or a question of a type the harness does not map (`score`), is Unsupported and counts as wrong.

On 35 counted benchmarks, 21 score exactly 0 skill (chance level), and the best is SGD/SGD-X at 8.84. Five benchmarks share a source with the training mixture (ANLI, HellaSwag, WinoGrande, GSM8K, CLINC150), yet score near chance: the model learned those datasets in the training mixture's question format and does not transfer to the index's formats. For example, CLINC150 scores 84.8% on this card's test but 0.19 skill on the index, which asks all 151 intents plus out-of-scope in one question. 140 training decisions share text with index test rows (74 HellaSwag, 26 MMLU, 25 MMLU-Pro, 10 ARC-Easy, 4 ARC-Challenge, 1 BBH).

<details>
<summary>Per-benchmark results (39 benchmarks)</summary>

| # | Benchmark | Skill (0 = chance, 100 = perfect) | Raw | Coverage | Unsupported | Counted in index |
| --- | --- | --- | --- | --- | --- | --- |
| #1 | BFCL | 2.94 | 0.2810 | 0.9799 | 34 | yes |
| #3 | API-Bank | 0.00 | 0.0000 | 0.0000 | 508 | yes |
| #4 | BANKING77 | 0.00 | 0.0014 | 1.0000 | 0 | yes |
| #5 | CLINC150+OOS | 0.19 | 0.0079 | 1.0000 | 0 | yes |
| #6 | RouterBench | 0.00 | 0.5304 | 0.9266 | 367 | yes |
| #9 | Home appliance simulator | 0.00 | 0.0000 | 1.0000 | 0 | yes |
| #10 | SGD/SGD-X | 8.84 | 0.4521 | 0.9996 | 1 | yes |
| #11 | ContractNLI | 0.00 | 0.0833 | 0.3659 | 78 | yes |
| #12 | ANLI | 0.30 | 0.3344 | 1.0000 | 0 | yes |
| #20 | BPoMP | 1.61 | 0.5081 | 1.0000 | 0 | yes |
| #21 | Humicroedit | 0.00 | 0.4920 | 1.0000 | 0 | yes |
| #23 | cfcolor | 0.00 | 0.4725 | 1.0000 | 0 | yes |
| #24 | MMLU | 0.73 | 0.2555 | 1.0000 | 0 | shown only |
| #25 | GPQA Diamond | 1.36 | 0.2602 | 1.0000 | 0 | yes |
| #26 | ARC-Easy | 2.39 | 0.2681 | 1.0000 | 0 | shown only |
| #27 | ARC-Challenge | 3.74 | 0.2782 | 1.0000 | 0 | shown only |
| #28 | WinoGrande | 0.72 | 0.5036 | 1.0000 | 0 | yes |
| #29 | HellaSwag | 3.25 | 0.2744 | 1.0000 | 0 | yes |
| #30 | GSM8K | 0.52 | 0.1778 | 1.0000 | 0 | yes |
| #32 | MuSR | 0.00 | 0.3577 | 1.0000 | 0 | yes |
| #33 | SATA-Bench | 0.00 | 0.0012 | 0.9594 | 67 | yes |
| #34 | SimpleBench | 0.00 | 0.0000 | 1.0000 | 0 | shown only |
| #37 | Amazon ESCI | 0.00 | 0.1965 | 1.0000 | 0 | yes |
| #38 | ACOS | 0.00 | 0.0000 | 1.0000 | 0 | yes |
| #39 | FinEntity | 0.00 | 0.2199 | 1.0000 | 0 | yes |
| #40 | iSarcasmEval | 5.16 | 0.2628 | 1.0000 | 0 | yes |
| #41 | VAST | 0.00 | 0.3276 | 1.0000 | 0 | yes |
| #42 | NLI4CT | 0.22 | 0.4871 | 0.9784 | 119 | yes |
| #43 | CRUXEval | 7.87 | 0.4193 | 1.0000 | 0 | yes |
| #44 | CLadder | 0.00 | 0.4976 | 1.0000 | 0 | yes |
| #48 | ForecastBench | 0.00 | 0.0000 | 0.9853 | 149 | yes |
| #50 | Habermas Machine | 0.00 | 0.2894 | 1.0000 | 0 | yes |
| #56 | PhishNChips phishing decisions | 0.00 | 0.4930 | 1.0000 | 0 | yes |
| #57 | MMLU-Pro | 0.02 | 0.1111 | 1.0000 | 0 | yes |
| #58 | BBH fixed-option tasks | 0.14 | 0.3111 | 1.0000 | 0 | yes |
| #59 | RAGTruth response-level hallucination | 0.00 | 0.2548 | 0.9837 | 44 | yes |
| #61 | HoVer claim verification | 0.00 | 0.4930 | 1.0000 | 0 | yes |
| #62 | When2Call MCQ | 0.00 | 0.2248 | 0.9721 | 102 | yes |
| #64 | New Yorker caption matching | 0.00 | 0.1932 | 1.0000 | 0 | yes |

</details>

<details>
<summary>Full per-task test results (58 tasks)</summary>

| Domain | Task | n | Accuracy | Chance | Held-out |
| --- | --- | --- | --- | --- | --- |
| agentic | `counsel/critique_quality` | 201 | 45.8% | 33.3% |  |
| agentic | `counsel/step_has_error` | 201 | 72.1% | 50.0% |  |
| agentic | `hotpotqa/comparison_yes_no` | 26 | 65.4% | 50.0% |  |
| agentic | `hotpotqa/retrieve` | 497 | 58.6% | 16.7% |  |
| classification | `ag_news/topic` | 500 | 84.6% | 25.0% |  |
| classification | `emotion/6way` | 500 | 11.6% | 16.7% | ✓ |
| classification | `sst5/score` | 500 | 27.4% | 20.0% | ✓ |
| classification | `yelp/score` | 500 | 48.6% | 20.0% |  |
| code | `bigclonebench/clone` | 500 | 95.4% | 50.0% |  |
| code | `codexglue/code_to_doc` | 504 | 83.9% | 26.2% |  |
| code | `codexglue/doc_to_code` | 504 | 88.5% | 27.9% |  |
| code | `codexglue/func_name` | 467 | 69.8% | 25.5% |  |
| code | `codexglue/lang_id` | 504 | 99.8% | 23.5% |  |
| code | `devign/vulnerability` | 500 | 64.2% | 50.0% |  |
| code | `humaneval/completion` | 119 | 42.0% | 39.4% | ✓ |
| code | `mbpp/bugspot` | 256 | 43.8% | 40.6% |  |
| code | `mbpp/solution` | 500 | 79.2% | 25.0% |  |
| guardrails | `agentharm/refuse` | 416 | 50.7% | 50.0% | ✓ |
| guardrails | `civil_comments/toxic` | 500 | 86.8% | 50.0% |  |
| guardrails | `jailbreak/detect` | 262 | 91.2% | 50.0% |  |
| guardrails | `prompt_injections/detect` | 116 | 47.4% | 50.0% | ✓ |
| intents | `banking77/intent` | 500 | 58.2% | 32.1% | ✓ |
| intents | `banking77/intent_77` | 300 | 4.7% | 1.3% | ✓ |
| intents | `clinc150/intent` | 500 | 84.8% | 12.4% |  |
| intents | `massive_en/intent` | 500 | 85.4% | 13.0% |  |
| policy | `policy/access_control_transfer` | 500 | 100.0% | 33.3% | ✓ |
| policy | `policy/count_threshold_transfer` | 500 | 34.8% | 17.8% | ✓ |
| policy | `policy/free_shipping_transfer` | 500 | 80.0% | 50.0% | ✓ |
| policy | `policy/invoice_overdue_transfer` | 500 | 48.8% | 50.0% | ✓ |
| policy | `policy/invoice_total_transfer` | 500 | 52.0% | 50.0% | ✓ |
| policy | `policy/refund_approval_transfer` | 500 | 52.0% | 33.3% | ✓ |
| policy | `policy/return_window_transfer` | 500 | 82.6% | 33.3% | ✓ |
| policy | `policy/sla_urgency_transfer` | 500 | 30.2% | 25.0% | ✓ |
| policy | `policy/table_extreme_transfer` | 500 | 28.4% | 23.7% | ✓ |
| reasoning | `anli/nli` | 498 | 32.1% | 33.3% |  |
| reasoning | `aqua_rat/math` | 247 | 21.9% | 20.0% |  |
| reasoning | `arc_challenge/mcq` | 500 | 29.4% | 25.0% | ✓ |
| reasoning | `arc_easy/mcq` | 500 | 30.4% | 25.0% | ✓ |
| reasoning | `boolq/yes_no` | 500 | 62.6% | 50.0% |  |
| reasoning | `commonsense_qa/mcq` | 493 | 19.1% | 20.0% |  |
| reasoning | `gsm8k/math` | 500 | 30.6% | 25.0% |  |
| reasoning | `hellaswag/continuation` | 500 | 28.0% | 25.0% |  |
| reasoning | `mmlu/mcq` | 500 | 28.4% | 25.0% | ✓ |
| reasoning | `mnli/claim` | 500 | 52.0% | 33.3% |  |
| reasoning | `openbookqa/mcq` | 500 | 36.2% | 25.0% |  |
| reasoning | `qasc/mcq` | 500 | 46.8% | 12.5% |  |
| reasoning | `sciq/mcq` | 498 | 60.8% | 25.0% |  |
| reasoning | `scitail/support` | 500 | 66.0% | 50.0% |  |
| reasoning | `snli/contradicts` | 500 | 85.4% | 33.3% |  |
| reasoning | `snli/must_be_true` | 500 | 90.8% | 33.3% |  |
| reasoning | `snli/nli` | 988 | 72.4% | 33.3% |  |
| reasoning | `winogrande/blank` | 500 | 56.2% | 50.0% |  |
| support | `bitext/category` | 500 | 99.6% | 16.1% |  |
| support | `bitext/route` | 500 | 99.4% | 19.0% |  |
| workflows | `typed_decisions/agent_trace_observability` | 500 | 69.8% | 30.0% |  |
| workflows | `typed_decisions/customer_service` | 500 | 50.0% | 28.0% |  |
| workflows | `typed_decisions/invoice_processing` | 500 | 56.0% | 35.0% |  |
| workflows | `typed_decisions/security_incidents` | 500 | 65.8% | 34.0% |  |

</details>

## Training

| | |
| --- | --- |
| Initialization | Random (trained from scratch; no pretrained weights, no tokenizer) |
| Data | 946,028 training / 26,597 test examples, plus a validation split from the same build |
| Sources | FalconDec_V2's mixture: public classification, NLI, QA, code, intent, guardrail and agent-trace datasets converted into typed decisions, plus synthetic policy-transfer and workflow tasks. `mind2web` was excluded; no custom data was added. |
| Preset | `full` (up to 40,000 training decisions per source) |
| Objective | Log score + 0.5 × spherical score, plus ranked probability score for `score` questions; option order reshuffled every epoch; "None of the above" augmentation on `choice` questions (8% replace, 4% insert); teacher probabilities mixed 50/50 with the answer where the data provides them; task sampling ∝ n^0.5, at most 3× repetition |
| Optimizer | AdamW, learning rate 1.5e-4, betas (0.9, 0.98), weight decay 0.01 (not on embeddings or norms), 6% warmup then cosine to 5%, gradient clip 1.0, EMA up to 0.999 |
| Batching | Length-sorted batches of up to 16,384 pooled positions and 64 decisions |
| Schedule | 20 epochs, 320,917 steps; EMA weights from **epoch 12** selected on validation macro accuracy |
| Calibration | Temperature per (question type × option-count bucket), fitted on validation by log loss |
| Compute | 121.6 minutes on one NVIDIA RTX PRO 6000 Blackwell Server Edition (training and test evaluation) |
| Software | PyTorch 2.9.0 (CUDA 13.0), Python 3.12 |


Validation accuracy peaked at epoch 12 and drifted down slightly afterwards, while many late training batches reached near-zero loss.

## Speed

Measured one decision at a time on 200 test decisions after 20 warm-up calls, including building the byte input.

| Device | Precision | Decisions per call | p50 | p90 |
| --- | --- | --- | --- | --- |
| cuda (RTX PRO 6000 Blackwell) | bf16 mixed precision (fp16 weights) | 1 | 5.14 ms | 5.26 ms |

GPU throughput: about **7,476 decisions per second** in batches of 128. CPU latency has not been measured yet.

## Intended use

LightDec_Arthur is built for high-volume, low-latency decisions where model size matters: ticket and email routing, intent detection, code and content screening, and pre-filtering before a more expensive model. At 24.5 MB it fits places a backbone-based model does not. Use the calibrated confidence to act on confident cases and send the rest to a person or a larger model.

## Limitations

- **Closed-set only.** The model can only choose among the options you supply. If the right answer is not listed, it will still pick something.
- **Weak outside its training formats.** Held-out tasks average 44.2% against 65.4% for tasks seen in training, and on the Decision Index it scores near chance, even on benchmarks whose source it trained on. Validate on your own questions before relying on it.
- **At or below chance on some tasks.** Emotion 11.6% (chance 16.7%, and 38.6% for always giving the most common label), ANLI 32.1% (chance 33.3%), CommonsenseQA 19.1% (20.0%), HellaSwag 28.0% (25.0%), MMLU 28.4% (25.0%), ARC-Challenge 29.4% (25.0%), prompt-injection detection 47.4% and harmful-request refusal 50.7% (50.0%). It is not a safety layer.
- **Fine-grained labels.** 77-way banking intents score 4.7%.
- **Policy transfer is uneven.** Scores range from 100% (`access_control`) to below the most-common-answer baseline (`invoice_overdue` 48.8% against 61.6%, `sla_urgency` 30.2% against 40.2%). At confidence ≥ 0.80 it acts on 41% of policy decisions but is right on only 77.4% of those.
- **Workflow decisions never reach 0.80 confidence** (0 of 2,000 test items), so with that threshold they always defer.
- **Calibration is aggregate.** Overall ECE is 0.030; per-task calibration has not been measured.
- **English only.** Questions are cut to 96 tokens (384 bytes) and each option to 24 tokens (96 bytes) by default.
- **No standalone loader yet**; see Usage.

## Files

| File | Description |
| --- | --- |
| `model.safetensors` | fp16 weights (12,268,033 parameters) |
| `config.json` | Tier configuration, byte layout and calibration temperatures |
| `results_full.json` | Full test results: per-task accuracy, speed, DecisionLab-style components and calibration |

## Citation

```bibtex
@misc{falconsai_lightdec_arthur_2026,
  title        = {LightDec_Arthur: A Byte-Level, Calibrated Decision Model},
  author       = {Falcons.ai},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Falconsai/LightDec_Arthur}}
}
```

## Architecture

- Byte-level input: 260 ids (256 bytes plus `[PAD]`, `[OPT]`, `[SEP]`, `[CLS]`); no tokenizer
- Hashed 2/3/4-byte fragment embeddings: 65,536 buckets × 128 dimensions, projected to 512
- Depthwise convolution (kernel 5), then 4:1 pooling
- One transformer block (width 512, 8 heads, MLP 1,536, rotary positions) applied 6 times over the text and 2 times over the options
- Option scorer: 2-layer MLP, one logit per option
- 12,268,033 parameters, stored in fp16
- License: to be confirmed (training-data licence review pending)

---


This card is generated from the surgical record itself; the package's
`lineage.intoto.jsonl` is the signed source of truth (verify it free at
the Surgeon's public verifier or with the bundled `verify_attestation.py`).

## Architecture
- Identification: **NLP · Small Language Model (SLM)** (80% confidence)
- Source format: `safetensors` · Intended task: not declared
- `config.json`: synthesized from the anatomy (no source config.json); model_type omitted — no architecture name in the source (QA-F-126)
- Source license: apache-2.0
- Lineage chain: 1 surgery (no prior attestation reachable) · Falconsai/LightDec_Arthur
- Post-surgery totals: 12,268,033 parameters ·
  19 tensors
- Compute estimate: 0.024519 GFLOPs (comparison
  metric, not a measurement)

## Provenance & operations
- Parents: Falconsai/LightDec_Arthur/model.safetensors
- Operations performed: load×1
- Weight merges recorded: 0
- Quantized tensors (F32→F16): 0

## Surgery Log (ordered)
1. **load** — hub:Falconsai/LightDec_Arthur/model.safetensors (24.5 MB, safetensors)

## Validation
- Tissue imaging: not run
- Structural integrity is testable offline via the packaged
  `load_and_test.py`.

## Compliance note
The signed attestation + this card together document model composition,
modification history, and validation evidence — the record structure
technical-documentation obligations (e.g. EU AI Act Annex IV) ask for.
This is evidence, not legal advice.

---
*Operated with Model Surgeon — verify this package at https://surgeon.falcons.ai/verify*
*© 2026 FALCONS.AI — Model Surgeon record format. The model weights remain their owner's.*