feat(server): serve GLiNER2.5-Decide typed-decision models

README.md

@@ -46,7 +46,7 @@ One SIE cluster runs the inference behind a whole agent. Each task is a handful
 | **Search** | Embed, match, and rerank to retrieve the right context. | [`bge-m3`](packages/sie_server/models/BAAI__bge-m3.yaml), [`splade-v3`](packages/sie_server/models/naver__splade-v3.yaml), [`colbertv2`](packages/sie_server/models/colbert-ir__colbertv2.0.yaml), [`qwen3-reranker`](packages/sie_server/models/Qwen__Qwen3-Reranker-4B.yaml) |
 | **Document to markdown** | PDFs, Office files, and scans become clean markdown. | [`lightonocr`](packages/sie_server/models/lightonai__LightOnOCR-2-1B.yaml), [`glm-ocr`](packages/sie_server/models/zai-org__GLM-OCR.yaml), [`mineru`](packages/sie_server/models/opendatalab__MinerU2.5-Pro-2604-1.2B.yaml), [`paddleocr-vl`](packages/sie_server/models/PaddlePaddle__PaddleOCR-VL-1.5.yaml), [`docling`](packages/sie_server/models/docling.yaml) |
 | **Structured output** | Schema-valid JSON, extracted or generated. | [`gliner2`](packages/sie_server/models/fastino__gliner2-large-v1.yaml), [`gliner-relex`](packages/sie_server/models/knowledgator__gliner-relex-large-v1.0.yaml), [`gliformer`](packages/sie_server/models/knowledgator__gliformer-large-v1.yaml), [`nuner-zero`](packages/sie_server/models/numind__NuNER_Zero.yaml), [`qwen3.8-27b`](packages/sie_server/models/Qwen__Qwen3.8-27B-FP8.yaml), [`qwen3.6-27b`](packages/sie_server/models/Qwen__Qwen3.6-27B.yaml) |
-| **Decide** | Choice, yes/no, and score answers with probabilities to typed questions about a text or JSON state. | [`laya`](packages/sie_server/models/convaiinnovations__laya.yaml), [`laya-multilingual`](packages/sie_server/models/convaiinnovations__laya-multilingual.yaml), [`laya-typed-decisions`](packages/sie_server/models/convaiinnovations__laya-typed-decisions.yaml) |
+| **Decide** | Choice, yes/no, and score answers with probabilities to typed questions about a text or JSON state. | [`laya`](packages/sie_server/models/convaiinnovations__laya.yaml), [`laya-multilingual`](packages/sie_server/models/convaiinnovations__laya-multilingual.yaml), [`laya-typed-decisions`](packages/sie_server/models/convaiinnovations__laya-typed-decisions.yaml), [`gliner2.5-decide`](packages/sie_server/models/fastino__GLiNER2.5-Decide.yaml), [`gliner2.5-multi-decide`](packages/sie_server/models/fastino__GLiNER2.5-multi-Decide.yaml), [`gliner2.5-decide-1b`](packages/sie_server/models/fastino__GLiNER2.5-Decide-1B.yaml) |
 | **Classify** | Zero-shot labels, with several label groups answered in one call. The instruct models also follow a task instruction and few-shot examples. | [`gliclass-large-v3`](packages/sie_server/models/knowledgator__gliclass-large-v3.0.yaml), [`gliclass-instruct-large`](packages/sie_server/models/knowledgator__gliclass-instruct-large-v1.0.yaml), [`gliclass-multilang-mini`](packages/sie_server/models/knowledgator__gliclass-multilang-mini.yaml) |
 | **Guard content** | A safety verdict: Yes/No with the threshold set in the model config, or safe/unsafe and policy-label scores with the threshold chosen per request. | [`granite-guardian-2b`](packages/sie_server/models/ibm-granite__granite-guardian-3.0-2b.yaml), [`opir-multitask-large`](packages/sie_server/models/knowledgator__opir-multitask-large-v1.0.yaml), [`opir-edge`](packages/sie_server/models/knowledgator__opir-edge-v1.0.yaml) |
 | **Run the agent loop** | Plan steps and call tools with an open LLM, streaming included. | [`qwen3.8-27b`](packages/sie_server/models/Qwen__Qwen3.8-27B-FP8.yaml), [`qwen3.6-27b`](packages/sie_server/models/Qwen__Qwen3.6-27B.yaml) |
@@ -69,7 +69,7 @@ docker run --gpus all -p 8080:8080 \
   -v sie-hf-cache:/app/.cache/huggingface \
   ghcr.io/superlinked/sie-server:latest-cuda12-default
 
-# Linux, NVIDIA GPU: Transformers 5 OCR models (LightOnOCR and GLM-OCR)
+# Linux, NVIDIA GPU: Transformers 5 models (LightOnOCR, GLM-OCR, and the GLiNER2.5-Decide models)
 docker run --gpus all -p 8080:8080 \
   -v sie-hf-cache:/app/.cache/huggingface \
   ghcr.io/superlinked/sie-server:latest-cuda12-transformers5
@@ -87,7 +87,8 @@ docker run -p 8080:8080 \
 
 Docker images are bundle-specific so dependency-incompatible model families stay isolated. Use the
 `sglang-vision-extract` image for LightOnOCR, GLM-OCR, and PaddleOCR-VL, or the `transformers5` image for the
-`:transformers` profiles of LightOnOCR and GLM-OCR; the `default` image intentionally does not advertise them.
+`:transformers` profiles of LightOnOCR and GLM-OCR and for the GLiNER2.5-Decide models; the `default` image
+intentionally does not advertise them.
 
 ```bash
 # in a second terminal
@@ -182,6 +183,30 @@ print(result["data"]["department"])  # values are illustrative and rounded
 # {'type': 'choice', 'choice': 'billing', 'probabilities': {'billing': 0.987, 'technical': 0.013}, 'confidence': 0.9}
 ```
 
+The GLiNER2.5-Decide models (`fastino/GLiNER2.5-Decide` for English, `GLiNER2.5-multi-Decide`, and
+`GLiNER2.5-Decide-1B`), served by the `transformers5` image from step 1, take the same questions, GLiClass-style `options={"label_groups": {...}}`, or plain `labels`,
+and return every option's probability. They read all of a call's questions next to the document in one row per item:
+one forward pass answers them all, and each question's probabilities depend on the other questions sent with it. A
+`score` question is read as the ordinal labels `"0"` to `"k-1"`, each described by its criterion; a `noul` question as
+`"yes"`/`"no"`. `usage.input_tokens` counts the document tokens the model reads plus the questions' instructions and
+criteria text; question ids and label names are not counted.
+
+```python
+result = client.extract(
+    "fastino/GLiNER2.5-Decide",
+    Item(text="Guest in room 1408 says the AC has been out since yesterday and wants to move rooms tonight."),
+    output_schema={
+        "intent": {"type": "choice", "instructions": "What does the guest want?",
+                   "criteria": {"room_change": "move to another room", "maintenance": "fix something", "checkout": None}},
+        "needs_human": {"type": "noul", "instructions": "Must a person act on this?"},
+        "urgency": {"type": "score", "instructions": "How urgent is this?", "criteria": ["low", "normal", "high", "urgent"]},
+    },
+)
+print(result["data"]["intent"])  # values are illustrative and rounded
+# {'type': 'choice', 'choice': 'room_change',
+#  'probabilities': {'room_change': 0.871, 'maintenance': 0.085, 'checkout': 0.045}, 'confidence': 0.57}
+```
+
 Text generation runs on the GPU generation image; stop the first server, then start this one on the same port:
 
 ```bash

packages/sie_sdk/README.md

@@ -160,6 +160,45 @@ graphs by default: their probabilities differed from eager execution by up to
 turn graphs on. See the server README for each model's measurements and the
 memory graphs use.
 
+## Typed decisions
+
+Typed-decision models answer typed questions about each item and return a
+probability for every option. The Laya models and the GLiNER2.5-Decide models
+take the same question mapping as `output_schema`: `choice` (pick one of the
+criteria), `noul` (yes or no), and `score` (an ordinal scale, index 0 first).
+Answers come back in `data`, keyed by question id.
+
+```python
+result = client.extract(
+    "fastino/GLiNER2.5-Decide",
+    Item(text="This is the third time I have explained the same missing refund. Get me a person."),
+    output_schema={
+        "intent": {
+            "type": "choice",
+            "instructions": "What does the customer want?",
+            "criteria": {"refund_request": "money back", "cancel": None, "complaint": "unhappy with service"},
+        },
+        "handoff": {"type": "noul", "instructions": "Should a person take over?"},
+        "frustration": {"type": "score", "instructions": "How frustrated is the customer?",
+                        "criteria": ["calm", "annoyed", "angry"]},
+    },
+)
+answers = result["data"]
+print(answers["intent"]["choice"], answers["intent"]["probabilities"])
+# e.g. refund_request {'refund_request': 0.59, 'cancel': 0.07, 'complaint': 0.34}
+print(answers["handoff"]["answer"], answers["handoff"]["noul"])  # True 0.99 (the probability of yes)
+print(answers["frustration"]["score"])  # e.g. 1.13, the expected level from 0 to 2
+```
+
+`choice` and `score` answers carry `probabilities` and `confidence`
+(`1 - entropy / log(number of options)`); a `noul` answer carries `noul`, the
+probability of yes, `answer`, and `confidence` (`max(p, 1 - p)`). The
+GLiNER2.5-Decide models also accept `options={"label_groups": {...}}` with
+`"classification_type": "multi-label"` for independent per-label scores, and
+plain `labels`, which return every label in `classifications`. They read all
+of a call's questions in one row per item, so each question's probabilities
+depend on the other questions sent with it.
+
 ## Generation prompts and guard verdicts
 
 `generate` and `stream_generate` treat text-only prompts as raw continuation

packages/sie_server/README.md

@@ -26,10 +26,15 @@ pip install sie-server
   pip install sie-server "transformers<5"
   ```
 
-- **OCR-VLM bundle** (LightOnOCR, GLM-OCR) — requires `transformers` 5.x, and is served with `-b transformers5`:
+- **Transformers 5 bundle** (LightOnOCR, GLM-OCR, GLiGuard, and the GLiNER2.5-Decide models) — requires
+  `transformers` 5.x, and is served with `-b transformers5`. The GLiNER2.5-Decide models also need `gliner2`
+  2.x. `sie-server` itself asks for `gliner2<2`, which the default bundle's GLiNER2 models need, so pip
+  reports that conflict when the second command below installs 2.x; the transformers5 bundle's GLiNER2
+  models are verified on 2.0.0:
 
   ```bash
   pip install sie-server "transformers>=5,<6"
+  pip install "gliner2==2.0.0"  # only for the GLiNER2.5-Decide models
   sie-server serve -b transformers5
   ```
 
@@ -169,6 +174,38 @@ largest shape on `gliclass-large-v1.0`. On a GPU shared with other models,
 leave memory headroom, or enable graphs only where the model has the GPU to
 itself. Usage and billing do not change.
 
+### GLiNER2.5-Decide usage and limits
+
+The GLiNER2.5-Decide models (`fastino/GLiNER2.5-Decide`, `GLiNER2.5-multi-Decide`,
+`GLiNER2.5-Decide-1B`) run on `gliner2` 2.x, which the transformers5 bundle
+pins (the `transformers5` image, or a native install as described above). Each
+item is one encoder row: every question's (or label group's) name,
+instruction, and labels, then the document. One forward pass answers them all,
+so the questions of a request are not independent: adding or changing one can
+change another's probabilities and score. `usage.input_tokens` counts the
+document tokens the model reads plus the tokens of the instructions and label
+descriptions (criteria) sent with the item, as Laya and GLiClass count
+instructions and criteria. Question ids, group names, and label names are not
+counted, and an item that returns an error counts nothing.
+
+A request takes at most 64 questions or label groups, 64 options per question,
+and 1,024 options in total. Question ids and group names may have 128
+characters, labels 256, and each instruction or description 2,048, with 65,536
+characters in all. Strings that contain one of the model's prompt markers
+(`[L]`, `[P]`, `[DESCRIPTION]`, ...) are refused. The questions may take at most
+512 tokens, or half the model's window when that is less: 256 of
+`GLiNER2.5-Decide`'s 512 tokens, 512 of the others' 2,048. This bounds the
+uncounted question and label tokens read with each item; a request needing more
+fails with `INPUT_TOO_LONG`. The
+document is read up to the whole words that fit in the rest of the window; a
+word longer than 4,096 characters, text past 64 characters per token of the
+window, or 4 words per token of the window also ends what is read. Words are
+split as gliner2 splits them, in linear time. A conversation (a list state) is
+read from its newest turn back; a run of more than 4,096 characters without a
+space is read only in its last 4,096 characters, and reading stops there. An item none of whose words fits, or that does
+not fit whole with `options={"overflow_policy": "error"}`, returns a per-item
+`INPUT_TOO_LONG` error while the other items succeed.
+
 ## Configuration
 
 `sie-server` reads its config from `SIE_*` environment variables (Pydantic

packages/sie_server/bundles/default.yaml

@@ -76,7 +76,10 @@ deps:
   requests: ''
   # gliner, gliner_bi (0.2.26: first release that loads the relex checkpoints)
   gliner: '>=0.2.26,<1'
-  # gliner2
+  # gliner2. Held at 1.x: gliner2 2.0 changes gliner2-base/large-v1 outputs
+  # (every repeated NER mention is returned, and structured extraction is
+  # rescored). The GLiNER2.5-Decide models, which need 2.x, are served from
+  # the transformers5 bundle.
   gliner2: '>=1.3.1,<2'
   # glirel
   glirel: '>=1.0,<2'

packages/sie_server/bundles/transformers5.yaml

@@ -7,6 +7,7 @@ adapters:
 - sie_server.adapters.lighton_ocr.adapter
 - sie_server.adapters.glm_ocr
 - sie_server.adapters.gliner2.classification
+- sie_server.adapters.gliner2.decide
 - sie_server.adapters.st_sparse_vision.adapter
 deps:
   # Transformers 5.x for models requiring newer architecture support
@@ -30,5 +31,10 @@ deps:
   # unsatisfiable against that pin and broke `uv run` resolution.
   # Let transformers drive the version.
   pillow: ''
-  # gliguard-LLMGuardrails-300M uses GLiNER2 classification APIs.
-  gliner2: '>=1.3.1,<2'
+  # gliguard-LLMGuardrails-300M uses GLiNER2 classification APIs, and the
+  # GLiNER2.5-Decide models need gliner2 2.x (AutoExtractor, boundary
+  # architecture). Exact: the Decide adapter builds the task prompt with the
+  # package's processor. GLiGuard's outputs are unchanged from 1.3.2. gliner2
+  # 2.0 declares transformers<5 only for its `local` extra, which is not used;
+  # the multilingual and 1B Decide checkpoints were saved with transformers 5.
+  gliner2: '==2.0.0'

packages/sie_server/models/fastino__GLiNER2.5-Decide-1B.yaml

@@ -0,0 +1,38 @@
+# GLiNER2.5-Decide-1B: typed-decision classifier on the Ettin 1B encoder
+# (jhu-clsp/ettin-enc-from-dec-1b, ModernBERT; span architecture), loaded through
+# gliner2 2.x AutoExtractor. Served from the transformers5 bundle, which pins gliner2 2.x.
+#
+# Checkpoint caveat: the weights at this revision are the step-285000 checkpoint
+# (upload commit eac5ac1e). The 59.6% fast-decisions score on the model card was
+# measured on an earlier step-195000 checkpoint (stated in that revision's card);
+# the uploaded weights had not been re-evaluated.
+#
+# The encoder config stores its RoPE bases only in the transformers-5
+# rope_parameters form (160000 for full and sliding-window layers). The adapter
+# checks the bases the loaded encoder actually uses, and under transformers 4
+# writes them where that version reads them (it would otherwise run the
+# sliding-window layers at 10000).
+sie_id: fastino/GLiNER2.5-Decide-1B
+hf_id: fastino/GLiNER2.5-Decide-1B
+hf_revision: 52c94d3b698bf6d2619df9d898bdc1523ea3f1ca
+inputs:
+  text: true
+  image: false
+  audio: false
+  video: false
+tasks:
+  encode: null
+  score: null
+  extract: {}
+# Tokens per item: every question's prompt and labels, then the document. The
+# Ettin encoder config sets max_position_embeddings 7999 and the tokenizer
+# model_max_length 8192; 2048 tokens stays within both.
+max_sequence_length: 2048
+profiles:
+  default:
+    max_batch_tokens: 16384
+    compute_precision: float16
+    adapter_path: sie_server.adapters.gliner2.decide:GLiNER2DecideAdapter
+    adapter_options:
+      loadtime: {}
+      runtime: {}

packages/sie_server/models/fastino__GLiNER2.5-Decide.yaml

@@ -0,0 +1,28 @@
+# GLiNER2.5-Decide: English typed-decision classifier (DeBERTa-v3-large encoder,
+# span architecture), loaded through gliner2 2.x AutoExtractor. Served from the
+# transformers5 bundle, which pins gliner2 2.x.
+sie_id: fastino/GLiNER2.5-Decide
+hf_id: fastino/GLiNER2.5-Decide
+hf_revision: 7ee5da4c2415e32259bcdc0b1a7367c32ce8d6f6
+inputs:
+  text: true
+  image: false
+  audio: false
+  video: false
+tasks:
+  encode: null
+  score: null
+  extract: {}
+# Tokens per item: every question's prompt and labels, then the document. 512,
+# as for fastino/gliner2-large-v1, this model's base: its DeBERTa-v3-large encoder
+# was pretrained on 512 positions (max_position_embeddings), and neither the model
+# card nor the checkpoint config (max_len null) gives a longer length.
+max_sequence_length: 512
+profiles:
+  default:
+    max_batch_tokens: 16384
+    compute_precision: float16
+    adapter_path: sie_server.adapters.gliner2.decide:GLiNER2DecideAdapter
+    adapter_options:
+      loadtime: {}
+      runtime: {}

packages/sie_server/models/fastino__GLiNER2.5-multi-Decide.yaml

@@ -0,0 +1,27 @@
+# GLiNER2.5-multi-Decide: multilingual typed-decision classifier (mDeBERTa-v3-base
+# encoder, boundary architecture; trained with max_len 4096 words), loaded through
+# gliner2 2.x AutoExtractor. Served from the transformers5 bundle, which pins gliner2 2.x.
+sie_id: fastino/GLiNER2.5-multi-Decide
+hf_id: fastino/GLiNER2.5-multi-Decide
+hf_revision: 6bc1d43d201b0691e733626389af8c57eea3ea68
+inputs:
+  text: true
+  image: false
+  audio: false
+  video: false
+tasks:
+  encode: null
+  score: null
+  extract: {}
+# Tokens per item: every question's prompt and labels, then the document. The
+# checkpoint config sets max_len 4096 (gliner2 counts it in words), and the
+# tokenizer sets no length limit; 2048 tokens stays within both.
+max_sequence_length: 2048
+profiles:
+  default:
+    max_batch_tokens: 16384
+    compute_precision: float16
+    adapter_path: sie_server.adapters.gliner2.decide:GLiNER2DecideAdapter
+    adapter_options:
+      loadtime: {}
+      runtime: {}

packages/sie_server/pyproject.toml

@@ -46,6 +46,8 @@ dependencies = [
     # never shipped, killing three extract cells at model load (#2873).
     # 0.2.26 is the first release that loads the joint entity-relation checkpoints.
     "gliner>=0.2.26,<1",
+    # gliner2 stays on 1.x with the default bundle (2.0 changes gliner2-base/large
+    # outputs); the transformers5 bundle overlays the 2.x pin its GLiNER2.5 models need.
     "gliner2>=1.3.1,<2",
     "glirel>=1.0,<2",
     # 0.1.17 adds the cross-attention scorer and pass-through pooling used by