Update model card (v4)

README.md

@@ -15,32 +15,17 @@ tags:
   - decode-free
 ---
 
-
-
-
 # Standard One 3B SH
 
+> **Updated weights (v4, 2026-10-22).** v4 is trained further on decision data, mostly our own synthetic items, with earlier data replayed. Decision Index 0.3 public index 58.08 (self-scored with the official kit; v3 51.53).
+> Earlier releases stay available under the tags `v3`, `v2` and `v1`.
 
-
-
-> **Updated weights (v3, 2026-10-13).** v3 is trained further on decision data, mostly our own synthetic items. Decision Index 0.3 public index 51.53 (self-scored with the official kit; v2 42.90).
-> Earlier releases stay available under the tags `v2` and `v1`.
-
-
-
-
-**Version:** v3
-
-
-
+**Version:** v4
 
 **Standard One 3B SH** is Standard One 3B with a **joint schema head** (SH). It reads the scenario once and
 scores every option of every question in a single forward pass, then returns probabilities through the same
 `POST /v1/systemone` contract as Standard One. It does not generate text.
 
-
-
-
 | | |
 |---|---|
 | Backbone | Standard One 3B v2.2 (`mistral3`, Ministral 3 3B text + Pixtral vision tower), fine-tuned for the head and merged; weights in this repository |
@@ -48,127 +33,78 @@ scores every option of every question in a single forward pass, then returns pro
 | Question types | `choice` (2–255 options), `noul` (yes/no), `score` (ordinal scale) |
 | Context | up to 32,768 tokens per request |
 | Prompt format | the chat template with **no default system message** (see "Prompt format") |
-| Other formats | [StandardOne-3B-SH-FP8](https://huggingface.co/StandardThinking/StandardOne-3B-SH-FP8) · [StandardOne-3B-SH-GGUF](https://huggingface.co/StandardThinking/StandardOne-3B-SH-GGUF) (both v3) |
+| Other formats | [StandardOne-3B-SH-FP8](https://huggingface.co/StandardThinking/StandardOne-3B-SH-FP8) · [StandardOne-3B-SH-GGUF](https://huggingface.co/StandardThinking/StandardOne-3B-SH-GGUF) (both still v3 for now) |
 | Larger model | [StandardOne-8B-SH](https://huggingface.co/StandardThinking/StandardOne-8B-SH) |
 | License | Apache-2.0 |
 
-
-
-
 ## Why a schema head
 
-
-
-
 Standard One reads one answer letter per option, so a question is limited by the label alphabet and every
 question in a request needs its own pass over the scenario. The schema head instead attends over the hidden states
 of the whole request (the final layer and one intermediate layer), places every question and option in one joint
 graph, and returns a calibrated distribution for each question at once. In practice:
 
-
-
-
 - **Many options.** Intent and routing questions with 100–255 options keep nearly the accuracy of small option sets.
 - **Many questions per scenario.** All questions of a request share one pass over the scenario.
 - **Long scenarios.** Decisions buried in long documents are found far more reliably than with Standard One v2.2.
 
-
-
-
 ## Results
 
-
-
-
-| Benchmark | SH v2 | **SH v3 (this release)** |
-|---|---:|---:|
-| Decision Index 0.3, public index | 42.90 | **51.53** |
-| Areas: Knowledge / Language / Retrieval / Tools / Arts | 25.2 / 44.2 / 59.7 / 60.3 / 19.8 | 30.3 / 49.5 / 63.3 / 79.8 / 36.5 |
-| HoVer, multi-hop claim verification (DI) | 62.7 | 63.9 |
-| BRIGHT, reasoning-heavy retrieval over long documents, ~12k tokens per request (DI) | 25.7 | 23.9 |
-| Decision hidden in a long document (our set, 150 items) | 82.00 | 85.33 |
-| JevBench public: easy / original / hard (48 / 72 / 111 items) | 100.00 / 98.61 / 63.96 | 100.00 / 98.61 / 63.96 |
-
-
-
+| Benchmark | SH v2 | SH v3 | **SH v4 (this release)** |
+|---|---:|---:|---:|
+| Decision Index 0.3, public index | 42.90 | 51.53 | **58.08** |
+| Areas: Knowledge / Language / Retrieval / Tools / Arts | 25.2 / 44.2 / 59.7 / 60.3 / 19.8 | 30.3 / 49.5 / 63.3 / 79.8 / 36.5 | 36.3 / 57.5 / 67.0 / 85.7 / 47.5 |
+| HoVer, multi-hop claim verification (DI) | 62.7 | 63.9 | 60.1 |
+| BRIGHT, reasoning-heavy retrieval over long documents, ~12k tokens per request (DI) | 25.7 | 23.9 | 36.1 |
+| Decision hidden in a long document (our set, 150 items) | 82.00 | 85.33 | 84.67 |
+| JevBench public: easy / original / hard (48 / 72 / 111 items) | 100.00 / 98.61 / 63.96 | 100.00 / 98.61 / 63.96 | 100.00 / 95.83 / 63.96 |
 
 Decision Index rows (DI) are chance-corrected skill scores (0 = random guessing, 100 = perfect) from our own runs of the public
-0.3 suite with the official kit, all 140,178 scoreable requests; every benchmark of both runs is in [standardone-3b-sh-v2](https://huggingface.co/datasets/StandardThinking/StandardOne-SH-decision-index/tree/main/runs/standardone-3b-sh-v2)
-and [standardone-3b-sh-v3](https://huggingface.co/datasets/StandardThinking/StandardOne-SH-decision-index/tree/main/runs/standardone-3b-sh-v3). For some DI benchmarks the public train split or the benchmark's public
+0.3 suite with the official kit, all 140,178 scoreable requests; every benchmark of all three runs is in [standardone-3b-sh-v2](https://huggingface.co/datasets/StandardThinking/StandardOne-SH-decision-index/tree/main/runs/standardone-3b-sh-v2),
+[standardone-3b-sh-v3](https://huggingface.co/datasets/StandardThinking/StandardOne-SH-decision-index/tree/main/runs/standardone-3b-sh-v3) and [standardone-3b-sh-v4](https://huggingface.co/datasets/StandardThinking/StandardOne-SH-decision-index/tree/main/runs/standardone-3b-sh-v4). For some DI benchmarks the public train split or the benchmark's public
 generator was part of training (HoVer among them); evaluated items never were. The other rows are accuracy in percent
-on identical items: the v2 column with the reference server in `code/`, the v3 column with the Standard One engine.
-
-
-
+on identical items: the v2 column with the reference server in `code/`, the v3 and v4 columns with the Standard One engine.
 
 ## Benchmark charts
 
-
-
-
 Decision Index 0.3 public index, board data of 2026-10-12; Standard One self-scored with the official kit (see the footnote of each chart).
 The same charts with notes: [Standard One SH on the Decision Index](https://standardthinking.ai/resources/guides/standard-one-sh-decision-index/).
 
-
-
-
 ![Public index against model size](https://huggingface.co/StandardThinking/StandardOne-3B-SH/resolve/main/assets/charts/01-index-vs-size.png)
 
-
-
-
 ![Retrieval & Classification](https://huggingface.co/StandardThinking/StandardOne-3B-SH/resolve/main/assets/charts/02-retrieval-classification.png)
 
-
-
-
 ![Tools & Automation](https://huggingface.co/StandardThinking/StandardOne-3B-SH/resolve/main/assets/charts/03-tools-automation.png)
 
-
-
-
 ![Product search relevance (Amazon ESCI)](https://huggingface.co/StandardThinking/StandardOne-3B-SH/resolve/main/assets/charts/04-product-search.png)
 
-
-
-
 ![Multi-hop claim verification (HoVer)](https://huggingface.co/StandardThinking/StandardOne-3B-SH/resolve/main/assets/charts/05-claim-verification.png)
 
-
-
-
 ![Smart-home request handling (Home appliance simulator)](https://huggingface.co/StandardThinking/StandardOne-3B-SH/resolve/main/assets/charts/06-home-appliance.png)
 
-
-
-
 ![Contract understanding (ContractNLI)](https://huggingface.co/StandardThinking/StandardOne-3B-SH/resolve/main/assets/charts/07-contract-nli.png)
 
-
-
-
 ![Reading chords from a score (POP909-CL)](https://huggingface.co/StandardThinking/StandardOne-3B-SH/resolve/main/assets/charts/08-chords.png)
 
+## Changes in v4
 
-
+- **Further training on decision data.** v4 continues from v3 with one more round of training on decision data, mostly our
+  own synthetic items, with earlier data replayed. The round also fixes answer-position and label-rate drift seen in the
+  previous checkpoint.
+- **What moved.** Compared with v3 (skill x100): PhishNChips 57.9 → 69.5, RAGTruth 21.9 → 50.6, GSM8K 17.2 → 41.2,
+  FinEntity 79.9 → 82.9. Down: POP909-CL 78.3 → 67.2, HoVer 63.9 → 60.1, GPQA Diamond 10.9 → 6.1.
+- **Internal and new-domain checks.** Slightly below the previous internal checkpoint: our own held-out new-domain set scored
+  65.9 against 67.1 for that checkpoint; JevBench original 95.83 (v3 98.61), hard unchanged at 63.96.
+- **Other formats.** The FP8 and GGUF repositories still hold v3 for now.
 
 ## Changes in v3
 
-
-
-
 - **Further training on synthetic decision data.** v3 continues from v2 on decision items built for this round, mostly our own
   synthetic items (code-checked, LLM-written in varied styles), with replay of earlier data. See "Results" for the
   Decision Index 0.3 and the long-context checks.
 
-
-
-
 ## Changes in v2
 
-
-
-
 - **Hard decision tasks.** v2 is fine-tuned on about 2,200 new decision items written in the families where v1 was
   weakest: long policies with exceptions and overrides, multi-criteria trade-offs, judging a response against a
   rubric, and date and number reasoning. The items were authored by large language models, every gold answer was
@@ -181,42 +117,24 @@ The same charts with notes: [Standard One SH on the Decision Index](https://stan
 - **Trade-offs.** Policy-judgement repair is 3.5 points below v1 (64.58 vs 68.08); WinoGrande, games and the intent
   label sets are within one point of v1. Keep `v1` if policy-judgement repair matters most to you.
 
-
-
-
 ## Prompt format
 
-
-
-
 The included `chat_template.jinja` has an empty default system message, and the reference server renders requests
 without one (`SH_SYSTEM_PROMPT=none`, the default in `code/render.py`). Do not add the Ministral default system
 prompt: the head was trained without it, and adding it changes the hidden states it reads.
 
-
-
-
 ## Quick start
 
-
-
-
 ```bash
 python -m pip install --upgrade huggingface_hub
 hf download StandardThinking/StandardOne-3B-SH --local-dir StandardOne-3B-SH
 python -m pip install -r StandardOne-3B-SH/code/requirements.txt
 
-
-
-
 python StandardOne-3B-SH/code/serve_head.py \
   --snapshot StandardOne-3B-SH --head StandardOne-3B-SH/head --port 30171 \
   --served-model-name standardthinking/standard-one-3b-sh
 ```
 
-
-
-
 ```bash
 curl -s localhost:30171/v1/systemone -H 'content-type: application/json' -d '{
   "state": {"message": "Please close my card, I lost it yesterday."},
@@ -229,25 +147,15 @@ curl -s localhost:30171/v1/systemone -H 'content-type: application/json' -d '{
 }'
 ```
 
-
-
-
 The server batches concurrent requests (`--max-batch-requests`, `--max-batch-tokens`) and binds `127.0.0.1` by
 default. It runs on one GPU with the backbone in bfloat16 and the head in float32. On Blackwell GPUs (B200/B300)
 install a CUDA 13 build of PyTorch first (`pip install torch --index-url https://download.pytorch.org/whl/cu130`).
 For CPU or Apple-silicon use, see [StandardOne-3B-SH-GGUF](https://huggingface.co/StandardThinking/StandardOne-3B-SH-GGUF).
 
-
-
-
-
-
 ## Run locally with llama.cpp
 
-
 For CPU or Apple-silicon use, follow the download, build and server instructions in [StandardOne-3B-SH-GGUF](https://huggingface.co/StandardThinking/StandardOne-3B-SH-GGUF#quick-start-with-llamacpp). The SH head reads hidden states from the backbone, so it requires the supplied state extractor and Python runner; stock `llama-server -hf` alone does not execute the schema head. Once the runner is listening on port 30171, use the requests below unchanged.
 
-
 ## Practical examples
 
 The following invented records illustrate four integrations: receiving inventory, diagnosing an import failure, evaluating an inspection note, and checking backup evidence. They are example inputs, not benchmark results. Start either the reference server or the SH GGUF runner on port 30171 before sending them.
@@ -368,9 +276,6 @@ These calls only request decisions; the surrounding application decides what to
 
 ## Files
 
-
-
-
 | Path | Contents |
 |---|---|
 | `model-0000{1,2}-of-00002.safetensors`, `config.json`, tokenizer files | merged backbone, bfloat16 (`Mistral3ForConditionalGeneration`) |
@@ -379,14 +284,8 @@ These calls only request decisions; the surrounding application decides what to
 | `code/` | reference `/v1/systemone` server: request rendering, head, scoring |
 | `SHA256SUMS`, `release-manifest.json` | checksums; source and merge details |
 
-
-
-
 ## Limitations
 
-
-
-
 - Requests longer than 32,768 tokens are refused, not truncated.
 - On our hard held-out decision suite (600 items; long policies, multi-step lookups, date and number reasoning)
   it scores 41.67 against 44.67 for Standard One 3B v2.2. Prefer Standard One 3B for that kind of workload.
@@ -394,16 +293,8 @@ These calls only request decisions; the surrounding application decides what to
 - The reference server is a PyTorch server; throughput under heavy load is lower than an inference engine.
 - Probabilities are calibrated on our held-out data; refit thresholds on your own distribution before relying on them.
 
-
-
-
 ## License
 
-
-
-
 Apache-2.0. Built on [Standard One 3B](https://huggingface.co/StandardThinking/StandardOne-3B) (Apache-2.0), which is
 built on [Ministral 3 3B](https://huggingface.co/mistralai/Ministral-3-3B-Instruct-2512) (Apache-2.0).
 
-
-

SHA256SUMS

@@ -1,32 +1,32 @@
 1495e1e757ef4d0925a2350563cf5754bb23c51701a8ec4fb3c5cdcbedae6747  LICENSE
-5b118d43ab6d82e9b6e3afd0837879b47a7e1f847195c5258a279651f61d5005  README.md
+c7cf4f5cf4ef2be0a95eb362d78f7005243c68de32d8b4161ea617f9b274608d  README.md
 331b249682cd52226c50e533f59825184997f3b31f9a62b2c3fea940db6999c5  SYSTEM_PROMPT.txt
+95a7b9bd0d618297650ab75632c7bf1f3d4ec1739119690d7c248bf3e9a00528  assets/charts/01-index-vs-size.png
+9b27154612b81237d02ac0384412cb1a41f9955e2b4db437f9aa25eb53b59f63  assets/charts/02-retrieval-classification.png
+a193f8032ef5d65c2783ecb6f0b43270b01abe95e4582b69dadfc285736888bd  assets/charts/03-tools-automation.png
+7f008ebd14e6d3074fb72a06f3a76f17734225bf02f897bb5b356dd0c1963e02  assets/charts/04-product-search.png
+a484d13ea51d3bd5a3501e808cf5447bf9cfefb2800cffd8410276dbbe3a1b2d  assets/charts/05-claim-verification.png
+10df32eb145fcd14176611c6229f034440d93683848ca28672677caa953936fa  assets/charts/06-home-appliance.png
+7f1c72e812c9652552bfcbc78210327b188435157296017b8b31cfc124432373  assets/charts/07-contract-nli.png
+47e9510297624fa13d5bba0deafd2f37441f33b23019d5364357af83274a5d5e  assets/charts/08-chords.png
 e62f05da25555a4c109d0f413eb673d328fcc40f7c82537c61631611f6954372  chat_template.jinja
+5f3d6035ca36885c9be04b7cfa88b5ba532c87d5b26fa3525c94327fb1b4348b  code/head.py
+88e3e1c200ff0d37f2a9fcb864ebf06b864e4883dcd71f2297d83bdf96a73120  code/model.py
+72ad6b7d1989105b8425c5d79e347fda2e9d3f2b429a5935c6d660983be412d1  code/render.py
+41685ab5516b7bf5eef3431b1fbcff43b35ed83bc957b88ffa4a4e23bf8ec855  code/requirements.txt
+9f18b699a0917d51b916e8ed96af87eb086af06fb9e6a723767032c107a50d15  code/scoring.py
+1b31b60dc0273ef671e230b3b1db79efd2a7de1410feef380757fc4cb2a316d9  code/serve_head.py
 3208d5b86b4ce169d39302f52944f8898e42c6917aab121524bd54cae5af9cb0  config.json
 e0923390059f84a9180b00e5501778acc45ea9856cd7f2fd68208b360927c677  generation_config.json
-7bbaf208fa3fb08ee9b398f6593240fc6fbca3d69532c6bcb95f1f5e100b4a79  model-00001-of-00002.safetensors
-6709c9c93807c19c89b598e14904ab7c491bc5cca0185f10f4edfb25a820aac6  model-00002-of-00002.safetensors
+d76a712bf7908d3d7d7d9f581817147ea8296bf2f4746db5022ba00936424448  head/schema_head.safetensors
+487be744adea0182f88f8a41c2a183226c29c7cd6bf8bb98a06ec9c2ff630b02  head/schema_head_config.json
+aca88073d0254fbd02507b8e0200f47cc254fb99aa3a7d4d6a3d08c918c50389  model-00001-of-00002.safetensors
+dbe1c06515b88895596fcad1745959681bbd1e29bf41137d9c2f115c2e563471  model-00002-of-00002.safetensors
 55641440eb385d8bc83a319b0a92f60fe45e6aef4b6992245eee91142237786e  model.safetensors.index.json
 d15587c57574d2a92c66ff2552c6db5173d4c440ad47f7454dc51bad0ecaeb59  params.json
 ece2373c2ae391bce18785a1810543bc6173a1f28b6767cffab2c35dbea5f002  processor_config.json
-a48416ec35f1e50f5add103b1902c6dbb8b1380aa9ee73ace4b9173ed89ea30f  release-manifest.json
+cfd7cf628d93586b239fe85cd3c11515ef0d16954e953b9a16419c191b73ba99  release-manifest.json
 0a5c981e8c5c6f8886ee007a6d4543a0be6b221cb9ca32a8709384a4c6fc8cbb  special_tokens_map.json
 600bb27946565481ecf51ba8aee252e49b9a68507866080ac9c30185bb312843  tekken.json
 d5f6046775b112f0e2d456ee9dba450684ab964fe5c4e231599bdc6773028135  tokenizer.json
 f59f7294e4f26383d0ea93840fe21cf197784be0842a8301a0343e8c34ed0d6d  tokenizer_config.json
-5f3d6035ca36885c9be04b7cfa88b5ba532c87d5b26fa3525c94327fb1b4348b  code/head.py
-88e3e1c200ff0d37f2a9fcb864ebf06b864e4883dcd71f2297d83bdf96a73120  code/model.py
-72ad6b7d1989105b8425c5d79e347fda2e9d3f2b429a5935c6d660983be412d1  code/render.py
-41685ab5516b7bf5eef3431b1fbcff43b35ed83bc957b88ffa4a4e23bf8ec855  code/requirements.txt
-9f18b699a0917d51b916e8ed96af87eb086af06fb9e6a723767032c107a50d15  code/scoring.py
-1b31b60dc0273ef671e230b3b1db79efd2a7de1410feef380757fc4cb2a316d9  code/serve_head.py
-86ab2a272b51d18fe7fffe78aab58bdf917d3b8ccf32b064efceeb5e19786aa7  head/schema_head.safetensors
-0dc3867d770916b77ebb7adfeab1aa0708a2b58a9428f390c84a2d0a71ff044e  head/schema_head_config.json
-95a7b9bd0d618297650ab75632c7bf1f3d4ec1739119690d7c248bf3e9a00528  assets/charts/01-index-vs-size.png
-9b27154612b81237d02ac0384412cb1a41f9955e2b4db437f9aa25eb53b59f63  assets/charts/02-retrieval-classification.png
-a193f8032ef5d65c2783ecb6f0b43270b01abe95e4582b69dadfc285736888bd  assets/charts/03-tools-automation.png
-7f008ebd14e6d3074fb72a06f3a76f17734225bf02f897bb5b356dd0c1963e02  assets/charts/04-product-search.png
-a484d13ea51d3bd5a3501e808cf5447bf9cfefb2800cffd8410276dbbe3a1b2d  assets/charts/05-claim-verification.png
-10df32eb145fcd14176611c6229f034440d93683848ca28672677caa953936fa  assets/charts/06-home-appliance.png
-7f1c72e812c9652552bfcbc78210327b188435157296017b8b31cfc124432373  assets/charts/07-contract-nli.png
-47e9510297624fa13d5bba0deafd2f37441f33b23019d5364357af83274a5d5e  assets/charts/08-chords.png

release-manifest.json

@@ -1,6 +1,6 @@
 {
- "version": "v3",
- "date": "2026-10-13",
+ "version": "v4",
+ "date": "2026-10-22",
  "backbone_base": {
   "repo": "StandardThinking/StandardOne-3B",
   "version": "v2.2"
@@ -15,11 +15,13 @@
    "hardgen-v1 hard decision items, soft gold 0.8, + replay and guard rows (136 steps)",
    "repair: policy/long-context/multi-hop + hardgen replay (57 steps)",
    "v2 (tag v2)",
-   "further training on decision data, mostly our own synthetic items, with replay (two rounds)"
+   "further training on decision data, mostly our own synthetic items, with replay (two rounds)",
+   "v3 (tag v3)",
+   "consolidation + L11/SATA round (m4e), drift-fix round (m4f), further training on decision data, mostly our own synthetic items, with replay (m4g)"
   ],
   "merged_sha256": {
-   "model-00001-of-00002.safetensors": "7bbaf208fa3fb08ee9b398f6593240fc6fbca3d69532c6bcb95f1f5e100b4a79",
-   "model-00002-of-00002.safetensors": "6709c9c93807c19c89b598e14904ab7c491bc5cca0185f10f4edfb25a820aac6"
+   "model-00001-of-00002.safetensors": "aca88073d0254fbd02507b8e0200f47cc254fb99aa3a7d4d6a3d08c918c50389",
+   "model-00002-of-00002.safetensors": "dbe1c06515b88895596fcad1745959681bbd1e29bf41137d9c2f115c2e563471"
   }
  },
  "head": {
@@ -29,24 +31,31 @@
   ],
   "template": "schema-v1-nosys",
   "parameters": 69513991,
-  "sha256": "86ab2a272b51d18fe7fffe78aab58bdf917d3b8ccf32b064efceeb5e19786aa7"
+  "sha256": "d76a712bf7908d3d7d7d9f581817147ea8296bf2f4746db5022ba00936424448"
  },
  "prompt_format": "chat template with an empty default system message (SH_SYSTEM_PROMPT=none)",
  "release": "StandardOne-3B-SH",
  "evaluation": {
-  "engine": "Standard One engine (SGLang fork ad26e8b179), 1 x AMD Instinct MI350X, 2026-10-13, released weights (agrees with production on 1,427 of 1,431 parity questions)",
+  "engine": "Standard One engine (ROCm build, image st-sglang:decision-qwen35-776e539961-60dac5b), 1 x AMD Instinct MI350X shared with production serving, 2026-10-21, released weights",
   "engine_clean_accuracy": {
-   "internal-eval-v1.heldout": 85.33,
-   "internal-eval-hardproxy-v1.heldout": 38.33,
-   "internal-eval-v2-realistic.heldout": 89.17,
+   "internal-eval-v1.heldout": 86.33,
+   "internal-eval-hardproxy-v1.heldout": 38.0,
+   "internal-eval-v2-realistic.heldout": 90.17,
    "jevbench-easy": 100.0,
-   "jevbench-original": 98.61,
+   "jevbench-original": 95.83,
    "jevbench-hard": 63.96,
-   "longctx-dev": 85.33,
-   "winogrande-dev": 91.59,
-   "repair5k-validation-policy-judgment": 57.0
+   "longctx-dev": 84.67,
+   "winogrande-dev": 92.74,
+   "repair5k-validation-policy-judgment": 58.83
   },
-  "decision_index_0.3_public_index": 51.53,
-  "decision_index_note": "self-scored with the official kit (apolinario/decision-index), all 140,178 scoreable requests"
+  "decision_index_0.3_public_index": 58.08,
+  "decision_index_note": "self-scored with the official kit (apolinario/decision-index @ 9eb2dbe, edition 0.3), all 140,178 scoreable requests; run runs/standardone-3b-sh-v4 of StandardThinking/StandardOne-SH-decision-index",
+  "decision_index_0.3_areas": {
+   "Knowledge & Reasoning": 36.26,
+   "Language Understanding": 57.53,
+   "Retrieval & Classification": 66.98,
+   "Tools & Automation": 85.74,
+   "Arts & Human Taste": 47.47
+  }
  }
-}
+}