dataset card (gate suite)

README.md

@@ -15,13 +15,35 @@ configs:
 
 # Gemma-4 Coder — tool-calling gate suite
 
-The **evaluation suite** behind the `tool-call pass rate` reported in our model cards'
-`model-index`. 8 cases: positive tool-use prompts (the model must emit a structured
-tool call) plus a no-tool **abstain** case (it must answer directly, not hallucinate a call).
+The 8-case eval behind the **tool-call pass rate** in our model cards' `model-index`:
+positive prompts where the model must emit a structured tool call, plus a no-tool
+**abstain** case (it must answer directly, not hallucinate a call). Use it to measure
+any local tool-calling model the same way we do — or to reproduce our numbers.
 
-Authored by us — fully permissive (apache-2.0). This is an *eval* set, not training data;
-our training mix is a derivative of public datasets and is **not** redistributed (its
-reproducible recipe — sources, pinned revisions, weights — is on each model card).
+```python
+from datasets import load_dataset
+cases = load_dataset("tpls/gemma4-coder-tool-eval", split="train")   # each: {name, user, tools, expect}
+```
+
+## Scoring a model
+
+Serve a GGUF on `llama.cpp` (`llama-server --jinja`), send each case's `user` + `tools`
+the normal OpenAI way, and check the reply against `expect`. **Two rates**, because
+`llama.cpp --jinja` doesn't recognise gemma-4's native tool-call markup:
+
+- **raw** — llama.cpp's native parse. For gemma-4 it *structurally undercounts* (blind to the format).
+- **shim** — the same outputs re-parsed for the native markup. This is the number that reflects training.
+
+The shim is a tiny serve-side post-processor — ready-to-use (drop-in litellm callback +
+a standalone parser, Apache-2.0) at **[`tpls/gemma4-tool-shim`](https://huggingface.co/tpls/gemma4-tool-shim)**,
+where the recovery algorithm is also documented so you can re-implement it.
+
+```python
+# sketch: per case, POST to your OpenAI-compatible endpoint, then
+#   raw_ok  = response had a structured tool_call (or correctly abstained)
+#   shim_ok = same, after running the reply through the shim parser
+# pass_rate = mean(ok over the 8 cases). expect == [] means "must NOT call a tool".
+```
 
 ## Files
 
@@ -30,19 +52,16 @@ reproducible recipe — sources, pinned revisions, weights — is on each model
 | `tool_eval_cases.jsonl` | one case/line: `{name, user, tools, expect}` |
 | `tool_calib.txt` | imatrix calibration text (code + tool-call markup) used when quantizing |
 
-## How it scores
-
-Serve a GGUF on `llama.cpp` (`llama-server --jinja`), send each case, and check whether a
-structured tool call was emitted (or correctly abstained). Two rates are reported:
-**raw** = llama.cpp's native parse; **shim** = the same output re-parsed for gemma-4's
-native markup (llama.cpp's `--jinja` doesn't recognise that format, so raw undercounts).
+## Scope & license
 
-```python
-from datasets import load_dataset
-cases = load_dataset("tpls/gemma4-coder-tool-eval", split="train")
-```
+Authored by us, fully permissive (apache-2.0). This is an **eval** set, not training data —
+our training mix is a derivative of public datasets and is **not** redistributed (its
+reproducible recipe lives on each model card). Pairs with the
+[`tpls/gemma4-tool-shim`](https://huggingface.co/tpls/gemma4-tool-shim) helper and the models below.
 
 ## Used by
 
+- [`tpls/Huihui-gemma-4-12B-coder-fable5-composer2.5-v1-abliterated-GGUF`](https://huggingface.co/tpls/Huihui-gemma-4-12B-coder-fable5-composer2.5-v1-abliterated-GGUF)
+- [`tpls/gemma-4-12B-coder-fable5-composer2.5-v1-abliterated-GGUF`](https://huggingface.co/tpls/gemma-4-12B-coder-fable5-composer2.5-v1-abliterated-GGUF)
 - [`tpls/gemma-4-12B-coder-fable5-composer2.5-v1-sft-v5`](https://huggingface.co/tpls/gemma-4-12B-coder-fable5-composer2.5-v1-sft-v5)
 - [`tpls/gemma-4-12B-coder-fable5-composer2.5-v1-sft-v5-GGUF`](https://huggingface.co/tpls/gemma-4-12B-coder-fable5-composer2.5-v1-sft-v5-GGUF)