Update model card

README.md

@@ -25,16 +25,14 @@ model-index:
 
 # karpo-decision-31B
 
-karpo-decision-31B is a decision model. Given a state (a web page, a document, a conversation: any text or JSON) and
-one or more typed questions about it, it returns a probability for every option of every question, read out of a
-single forward pass instead of generated text. It is fine-tuned from
-[google/gemma-4-31B-it](https://huggingface.co/google/gemma-4-31B-it) and adds a linear readout over the options.
+A decision model: given a state (a web page, a document, a conversation: any text or JSON) and typed questions about
+it, it returns a probability for every option of every question in one forward pass, without generating text.
+Fine-tuned from [google/gemma-4-31B-it](https://huggingface.co/google/gemma-4-31B-it), with a linear readout over the
+options.
 
-- **Base model:** [google/gemma-4-31B-it](https://huggingface.co/google/gemma-4-31B-it)
 - **Parameters:** 31.3B in bf16, plus a 255 × 5,376 readout in fp32
-- **Questions:** `choice` (pick one of up to 255 options), `noul` (true or false) and `score` (a level on a scale)
-- **Context:** 32,768 tokens per sequence; a longer request is split into parts of its questions, each with the full
-  state
+- **Questions:** `choice` (up to 255 options), `noul` (true or false), `score` (a level on a scale)
+- **Context:** 32,768 tokens per sequence; a longer request is split by questions, each part with the full state
 - **Licence:** Apache-2.0
 
 ## Lineage
@@ -44,8 +42,8 @@ single forward pass instead of generated text. It is fine-tuned from
 
 ## Use
 
-The inference code is [github.com/machinepulse-ai/karpo-decision](https://github.com/machinepulse-ai/karpo-decision)
-(Apache-2.0, Python 3.12+). It runs on one GPU: the weights take 62.5 GB, and long requests need more on top.
+Inference code: [github.com/machinepulse-ai/karpo-decision](https://github.com/machinepulse-ai/karpo-decision). One GPU;
+the weights take 62.5 GB, long requests more.
 
 ```bash
 pip install "karpo-decision @ git+https://github.com/machinepulse-ai/karpo-decision"
@@ -71,13 +69,13 @@ print(response["answers"])
 #  "action": {"type": "choice", "choice": <key>, "probabilities": {<key>: p, ...}, "confidence": c}}
 ```
 
-As an HTTP server (`POST /v1/systemone`, `GET /healthz`):
+HTTP server (`POST /v1/systemone`, `GET /healthz`):
 
 ```bash
 torchrun --standalone --nproc_per_node 1 -m karpo_decision.serve.server --checkpoint "$path" --port 18440
 ```
 
-On the [Decision Index](https://github.com/apolinario/decision-index) kit (installed separately):
+[Decision Index](https://github.com/apolinario/decision-index) kit:
 
 ```bash
 python -m decision_index run --engine karpo_decision.serve.engine:DecisionIndexEngine \
@@ -86,9 +84,8 @@ python -m decision_index run --engine karpo_decision.serve.engine:DecisionIndexE
 
 ## Results
 
-Decision Index 0.3, public suite, run by us on the kit's suite and scored with the kit: 140,178 requests, 140,167
-answered and 11 unsupported (a question that did not fit in 32,768 tokens with its state), default reading
-(`view=joint`), torch 2.14.1, transformers 5.21.0.
+Decision Index 0.3 public suite, self-reported: 140,178 requests, 11 unsupported (a question over 32,768 tokens with
+its state).
 
 | | score |
 |---|---|
@@ -101,21 +98,18 @@ answered and 11 unsupported (a question that did not fit in 32,768 tokens with i
 | Tools & Automation | 88.32 |
 | Arts & Human Taste | 68.10 |
 
-The leaderboard ranks models by the Full score, which its maintainers compute on private tests; we will add it here
-once it is published.
+The Full score (the leaderboard's private tests) will be added once published.
 
 ## Limitations
 
-- The released inference code reads text only. The vision weights are in the checkpoint but unused.
-- It is not a chat or text-generation model: it was trained to answer through the readout, and generation was not
-  evaluated.
-- Probabilities use temperature 1.0 (`decision_config.json`); they are not calibrated after training.
-- A question that does not fit in 32,768 tokens together with the state is refused (`Unsupported`), never truncated.
-- It always picks among the given options; it abstains only when an option says so.
+- Text only: the vision weights are in the checkpoint but unused.
+- Not a chat or text-generation model; generation was not evaluated.
+- Probabilities are not calibrated (temperature 1.0).
+- A question that does not fit in 32,768 tokens with its state is refused, never truncated.
+- It always picks one of the given options.
 
 ## Licence
 
-The weights are released under the Apache License 2.0 ([`LICENSE`](LICENSE)), as is the inference code. They are a
-modified version of [google/gemma-4-31B-it](https://huggingface.co/google/gemma-4-31B-it), which Google releases under
-the [Apache License 2.0](https://ai.google.dev/gemma/docs/gemma_4_license): its decoder weights were fine-tuned and a
-readout was added.
+Apache-2.0 ([LICENSE](LICENSE)), for the weights and the inference code. A modified version of
+[google/gemma-4-31B-it](https://huggingface.co/google/gemma-4-31B-it)
+([Apache-2.0](https://ai.google.dev/gemma/docs/gemma_4_license)): decoder weights fine-tuned, readout added.