v3: further training on synthetic decision data (Decision Index 0.3 public index, self-scored)

README.md

@@ -17,10 +17,10 @@ tags:
 
 # Standard One 3B SH
 
-> **Updated weights (v2, 2026-10-11).** v2 is trained further on hard decision tasks (long policies, trade-offs,
-> judging, dates and numbers). The previous release stays available under the tag `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:** v2
+**Version:** v3
 
 **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
@@ -33,7 +33,7 @@ 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) |
+| 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) (being updated to v3) |
 | Larger model | [StandardOne-8B-SH](https://huggingface.co/StandardThinking/StandardOne-8B-SH) |
 | License | Apache-2.0 |
 
@@ -68,7 +68,13 @@ All numbers compare models on identical items, scored the same way.
 | Held-out realistic decision suite (600 items) | 88.17 | **88.50** | 88.33 |
 | Held-out hard decision suite (600 items) | **44.67** | 40.17 | 41.67 |
 
-Accuracy in percent. The Decision Index for v2 is being measured and will be added here.
+Accuracy in percent. On the Decision Index 0.3 public suite (self-scored with the official kit) v2 scores 42.90 and v3 51.53.
+
+## 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. The results table above
+  describes v2; Decision Index 0.3: 51.53 (v2: 42.90).
 
 ## Changes in v2
 
@@ -118,51 +124,6 @@ default. It runs on one GPU with the backbone in bfloat16 and the head in float3
 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).
 
-## Fast serving with the Standard One engine
-
-The reference server above is plain PyTorch. For production throughput and latency, serve this repository with the
-**Standard One engine**, our SGLang fork with a prefill-only decision mode and the schema-head readout
-([standard-one-sglang](https://github.com/standardthinkingai/standard-one-sglang),
-[standard-one-adapter](https://github.com/standardthinkingai/standard-one-adapter), Apache-2.0). It is the engine we run
-in production. Build the image once (it replaces only the Python package of a stock SGLang image):
-
-```bash
-mkdir ctx && cd ctx
-git clone https://github.com/standardthinkingai/standard-one-sglang sglang-jev
-git clone https://github.com/standardthinkingai/standard-one-adapter jev-adapter
-mkdir -p jit-cache/aiter jit-cache/root-cache
-docker build -f sglang-jev/examples/runtime/jev/rocm/Dockerfile.decision \
-  --build-arg BASE_IMAGE=lmsysorg/sglang:nightly-dev-cu13-20260929-79cafec0 -t standard-one-engine:cu13 .
-```
-
-Then serve the downloaded repository as it is (weights at the top level, head in `head/`):
-
-```bash
-docker run -d --gpus all --network host --ipc host --shm-size 32g \
-  -v $PWD/StandardOne-3B-SH:/models/m:ro standard-one-engine:cu13 python3 -m sglang.launch_server \
-    --model-path /models/m --decision-readout-path /models/m/head --served-model-name standardthinking/standard-one-3b-sh \
-    --host 127.0.0.1 --port 30000 --tp-size 1 --model-impl sglang --dtype bfloat16 --attention-backend triton \
-    --context-length 32768 --max-running-requests 64 --mem-fraction-static 0.3 \
-    --chunked-prefill-size -1 --model-config-parser hf --load-format safetensors \
-    --disable-radix-cache --disable-decode-cuda-graph --disable-prefill-cuda-graph --decision-deadline-seconds 110
-```
-
-It answers `POST /v1/systemone` on port 30000. Unlike the reference server, the engine requires the `model` field
-(the served model name), as in the System One API:
-
-```bash
-curl -s localhost:30000/v1/systemone -H 'content-type: application/json' -d '{
-  "model": "standardthinking/standard-one-3b-sh",
-  "state": {"message": "Please close my card, I lost it yesterday."},
-  "questions": {"urgent": {"type": "noul", "instructions": "The request is urgent."}}
-}'
-```
-
- Checked on one NVIDIA B300 against the reference
-server on 966 evaluation questions: 98.96 % the same answer as the reference server; one small request 24 ms median, one request at a time. On AMD MI350X/MI355X build with
-`BASE_IMAGE=lmsysorg/sglang-rocm:v0.5.20-rocm720-mi35x-20260929` and run with
-`--device /dev/kfd --device /dev/dri --group-add video -e SGLANG_USE_AITER=0` instead of `--gpus all`.
-
 ## Files
 
 | Path | Contents |
@@ -179,7 +140,7 @@ server on 966 evaluation questions: 98.96 % the same answer as the reference ser
 - 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.
 - Text input only has been validated for this release.
-- The reference server is a PyTorch server; for throughput use the Standard One engine (see above).
+- 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

SHA256SUMS

@@ -1,24 +1,24 @@
 1495e1e757ef4d0925a2350563cf5754bb23c51701a8ec4fb3c5cdcbedae6747  LICENSE
-38a5fb1e614034d82a1ddc483951d5485657c93bcd828c2452bf489d4f6fb4f3  README.md
+a884648191a35b9967e24140e697603d0d186f254fa5976c5ef59f3ead5c6489  README.md
 331b249682cd52226c50e533f59825184997f3b31f9a62b2c3fea940db6999c5  SYSTEM_PROMPT.txt
 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
-98eca2c44d1d47626d03e8f5f93974f9e92330b03751489e313f350526ec1936  head/schema_head.safetensors
-cc27869adb938796baadbeb70f8bbafb5c2309748d47ba9cf6908ec9d976ab96  head/schema_head_config.json
-f2a51d96a99dd3dcb85330542c7ae069ec11eb7cf01124af053a7a0ba14d67f9  model-00001-of-00002.safetensors
-d2d8992864001000f98180c515b77fb73a1f6336b26a24b28557922f30768a2e  model-00002-of-00002.safetensors
+7bbaf208fa3fb08ee9b398f6593240fc6fbca3d69532c6bcb95f1f5e100b4a79  model-00001-of-00002.safetensors
+6709c9c93807c19c89b598e14904ab7c491bc5cca0185f10f4edfb25a820aac6  model-00002-of-00002.safetensors
 55641440eb385d8bc83a319b0a92f60fe45e6aef4b6992245eee91142237786e  model.safetensors.index.json
 d15587c57574d2a92c66ff2552c6db5173d4c440ad47f7454dc51bad0ecaeb59  params.json
 ece2373c2ae391bce18785a1810543bc6173a1f28b6767cffab2c35dbea5f002  processor_config.json
-6bee24779475454b58b56cbb956e6c35d7053ce16cc94a8347e10bbcff606c71  release-manifest.json
+535a89c61b8551783e056c09f5111df116823b2ea107ca18d260d64e78e997ee  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

head/schema_head.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:98eca2c44d1d47626d03e8f5f93974f9e92330b03751489e313f350526ec1936
+oid sha256:86ab2a272b51d18fe7fffe78aab58bdf917d3b8ccf32b064efceeb5e19786aa7
 size 278062780

head/schema_head_config.json

@@ -1,30 +1,30 @@
 {
-  "head": {
-    "hidden_size": 3072,
-    "n_taps": 2,
-    "d": 1024,
-    "heads": 16,
-    "ffn": 2816,
-    "route_blocks": 2,
-    "interact_blocks": 2,
-    "dropout": 0.1,
-    "n_roles": 5,
-    "ord_hidden": 256,
-    "max_nodes": 4096,
-    "grad_checkpoint": true,
-    "taps": [
-      "final",
-      19
-    ],
-    "template": "schema-v1"
-  },
+ "head": {
+  "hidden_size": 3072,
+  "n_taps": 2,
+  "d": 1024,
+  "heads": 16,
+  "ffn": 2816,
+  "route_blocks": 2,
+  "interact_blocks": 2,
+  "dropout": 0.1,
+  "n_roles": 5,
+  "ord_hidden": 256,
+  "max_nodes": 4096,
+  "grad_checkpoint": true,
   "taps": [
-    "final",
-    19
+   "final",
+   19
   ],
-  "template": "schema-v1-nosys",
-  "step": 57,
-  "lora_rank": 64,
-  "snapshot": "/scratch/schema-head/backbones/v22-3b",
-  "preview": true
-}
+  "template": "schema-v1"
+ },
+ "taps": [
+  "final",
+  19
+ ],
+ "template": "schema-v1-nosys",
+ "step": 1113,
+ "lora_rank": 64,
+ "snapshot": "/scratch/schema-head/models/StandardOne-3B-SH",
+ "preview": true
+}

model-00001-of-00002.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f2a51d96a99dd3dcb85330542c7ae069ec11eb7cf01124af053a7a0ba14d67f9
+oid sha256:7bbaf208fa3fb08ee9b398f6593240fc6fbca3d69532c6bcb95f1f5e100b4a79
 size 4995664864

model-00002-of-00002.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d2d8992864001000f98180c515b77fb73a1f6336b26a24b28557922f30768a2e
+oid sha256:6709c9c93807c19c89b598e14904ab7c491bc5cca0185f10f4edfb25a820aac6
 size 2702576200

release-manifest.json

@@ -1,6 +1,6 @@
 {
- "version": "v2",
- "date": "2026-10-11",
+ "version": "v3",
+ "date": "2026-10-13",
  "backbone_base": {
   "repo": "StandardThinking/StandardOne-3B",
   "version": "v2.2"
@@ -13,11 +13,13 @@
    "v1 (3B SH stage A)",
    "P7 teacher-label continuation (2,077 steps)",
    "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)"
+   "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)"
   ],
   "merged_sha256": {
-   "model-00001-of-00002.safetensors": "f2a51d96a99dd3dcb85330542c7ae069ec11eb7cf01124af053a7a0ba14d67f9",
-   "model-00002-of-00002.safetensors": "d2d8992864001000f98180c515b77fb73a1f6336b26a24b28557922f30768a2e"
+   "model-00001-of-00002.safetensors": "7bbaf208fa3fb08ee9b398f6593240fc6fbca3d69532c6bcb95f1f5e100b4a79",
+   "model-00002-of-00002.safetensors": "6709c9c93807c19c89b598e14904ab7c491bc5cca0185f10f4edfb25a820aac6"
   }
  },
  "head": {
@@ -27,21 +29,24 @@
   ],
   "template": "schema-v1-nosys",
   "parameters": 69513991,
-  "sha256": "98eca2c44d1d47626d03e8f5f93974f9e92330b03751489e313f350526ec1936"
+  "sha256": "86ab2a272b51d18fe7fffe78aab58bdf917d3b8ccf32b064efceeb5e19786aa7"
  },
  "prompt_format": "chat template with an empty default system message (SH_SYSTEM_PROMPT=none)",
  "release": "StandardOne-3B-SH",
  "evaluation": {
-  "reference_server": {
-   "internal-eval-v1.heldout": 84.33,
-   "internal-eval-hardproxy-v1.heldout": 41.67,
-   "internal-eval-v2-realistic.heldout": 88.33,
+  "engine": "Standard One engine (SGLang fork ad26e8b179), 1 x AMD Instinct MI350X, 2026-10-13",
+  "engine_clean_accuracy": {
+   "internal-eval-v1.heldout": 84.0,
+   "internal-eval-hardproxy-v1.heldout": 38.83,
+   "internal-eval-v2-realistic.heldout": 89.17,
    "jevbench-easy": 100.0,
    "jevbench-original": 98.61,
-   "jevbench-hard": 63.96,
-   "longctx-dev": 82.0,
-   "winogrande-dev": 92.19,
-   "repair5k-validation-policy-judgment": 53.33
-  }
+   "jevbench-hard": 60.36,
+   "longctx-dev": 88.0,
+   "winogrande-dev": 91.19,
+   "repair5k-validation-policy-judgment": 56.5
+  },
+  "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"
  }
 }