Instructions to use aimeri/spoomplesmaxx-whiskeyjack-12B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use aimeri/spoomplesmaxx-whiskeyjack-12B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="aimeri/spoomplesmaxx-whiskeyjack-12B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("aimeri/spoomplesmaxx-whiskeyjack-12B")
model = AutoModelForMultimodalLM.from_pretrained("aimeri/spoomplesmaxx-whiskeyjack-12B", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use aimeri/spoomplesmaxx-whiskeyjack-12B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "aimeri/spoomplesmaxx-whiskeyjack-12B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aimeri/spoomplesmaxx-whiskeyjack-12B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/aimeri/spoomplesmaxx-whiskeyjack-12B

SGLang

How to use aimeri/spoomplesmaxx-whiskeyjack-12B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "aimeri/spoomplesmaxx-whiskeyjack-12B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aimeri/spoomplesmaxx-whiskeyjack-12B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "aimeri/spoomplesmaxx-whiskeyjack-12B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aimeri/spoomplesmaxx-whiskeyjack-12B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use aimeri/spoomplesmaxx-whiskeyjack-12B with Docker Model Runner:
```
docker model run hf.co/aimeri/spoomplesmaxx-whiskeyjack-12B
```

aimeri commited on Jul 30

Commit

9828408

verified ·

1 Parent(s): 398d5da

Update README.md

Browse files

Files changed (1) hide show

README.md +480 -87

README.md CHANGED Viewed

@@ -1,101 +1,494 @@
 ---
 license: gemma
-base_model: google/gemma-4-12B
 library_name: transformers
 pipeline_tag: text-generation
-tags: [gemma4, sft, reasoning, roleplay, tool-use]
 ---
-# spoomplesmaxx-whiskeyjack-12B
-gemma-4-12B full-parameter SFT (ms-swift `swift sft`, DeepSpeed ZeRO-2, torch SDPA attention, custom liger fused-CE).
-Control tokens (`<turn|>`, `<|channel>`/`<channel|>`, `<|tool_call>`/`<tool_call|>`)
-were audited pre-training and verified emit-able post-training (stop battery,
-boundary probes, tool-call battery).
-## Format
-Gemma 4 turn format — note it differs from Gemma 3 entirely (there is no
-`<start_of_turn>`), and the assistant role is spelled `model`:
-```
-<|turn>user
-...<turn|>
-<|turn>model
-<|channel>thought
 ...reasoning...
-<channel|>...answer...<turn|>
-```
-- **Stops on** `<turn|>` (106). `<eos>` (1) is kept as a secondary EOS.
-  Do not configure `<eos>` alone as the stop — turns end on `<turn|>`.
-- **Thinking** is selected by a `<|think|>` marker at the top of the **system**
-  turn, which `apply_chat_template(enable_thinking=True)` injects. It defaults to
-  **False**. Both modes share the bare `<|turn>model` generation prefix; the model
-  decides whether to open `<|channel>thought` itself.
-- **The chat template here is NOT stock Gemma 4.** Upstream appends an empty
-  channel (`<|channel>thought
-<channel|>`) for non-thinking. That form appears
-  **0 times in 1000 training turns** of this corpus — a no-thoughts turn carries no
-  channel at all — so it is deliberately removed. The original is kept alongside as
-  `chat_template.gemma-it-original.jinja`. Restoring it puts the model out of
-  distribution (reasoning leaks into content, tool calls lose their opener).
-- **Tool calls** use Gemma's DSL, not JSON:
-  `<|tool_call>call:NAME{key:<|"|>value<|"|>}<tool_call|>`.
-- **Serving tool calls — read this or you will see "infinite tool loops."** In the
-  training corpus a whole tool episode lives inside ONE `<|turn>model`, with
-  `<|tool_response>` blocks interleaved *inline*; the model never emitted `<turn|>`
-  after a call and so never learned to yield. Serve with **`stop=["<tool_call|>"]`**,
-  inject the result as `<|tool_response>response:NAME{...}<tool_response|>`, and
-  continue the SAME turn. A harness that waits for `<turn|>` after a call will
-  hang and the model will keep generating plausible calls.
-## Training stages
-| stage | data | steps | LR | eval_loss | token_acc |
-|---|---|---|---|---|---|
-| 1 (tag `v1-baseline-rp`) | aviary burn corpus, 1 epoch | 1,917 | 1e-5 | 1.311 | 0.6465 |
-| 2 (tag `v2-corrected-rp`) | thinking-weighted resample | 568 | 2e-6 | 1.3067 | 0.6479 |
-| 3 (**main**) | + 4,000 converted RP-reasoning rows | 574 | 2e-6 | **1.301** | **0.6491** |
-Stage 3 exists because stage 2 thinking was entangled with a single system
-framing: 19,605 of the corpus's 20,666 thought-bearing rows share one RP prompt
-shape, so the model opened a thought channel on 8/8 in-corpus rows and 1/25 real
-character cards. Stage 3 interleaved RP-reasoning rows under ~4,000 distinct
-character cards to decouple the two.
-## Measured behaviour
-On 25 held-out character cards (846-5,053 chars) with a short opening message:
-| | stage 2 | stage 3 |
-|---|---|---|
-| opens a thought channel | 1/25 (4%) | **23/25 (92%)** |
-Unchanged across the pass: stop rate 10/10 in both thinking and non-thinking
-modes, tool-call round trip, 0/3 stray channels when thinking is off, and
-P(`<channel|>`) at the true close = 1.000.
-### Two behaviours to expect
-**The thought form depends on the system prompt shape.** Under a SillyTavern-style
-character card the model writes a structured planner
-(`SCENE:` / `CHARACTERS:` / `CONTINUITY:` / `THREADS:` / `PLAN:`, ~750 chars,
-23/23 of the cards that opened). Under the corpus's own RP framing it writes
-short first-person interiority (~90 chars). These are two distinct learned forms
-conditioned on the prompt, not a blend.
-**Thinking is optional under a companion persona.** A conversational companion
-card produces no thought channel (0/6) — that reflects the training corpus, where
-such rows are predominantly non-thinking. This is by design, not a defect.
-- **generation_config:** temperature 1.0, top_p 0.95,
-  top_k 64.
-```python
 from transformers import AutoModelForImageTextToText, AutoTokenizer
 tok = AutoTokenizer.from_pretrained("aimeri/spoomplesmaxx-whiskeyjack-12B")
-model = AutoModelForImageTextToText.from_pretrained("aimeri/spoomplesmaxx-whiskeyjack-12B", dtype="bfloat16", device_map="auto")
 msgs = [{"role": "user", "content": "Solve (x + 2)^2 = 0."}]
-ids = tok.apply_chat_template(msgs, add_generation_prompt=True, enable_thinking=True,
-                              return_tensors="pt").to(model.device)
 out = model.generate(ids, max_new_tokens=512)
 print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=False))
-```

 ---
 license: gemma
+base_model:
+- google/gemma-4-12B
 library_name: transformers
 pipeline_tag: text-generation
+tags:
+- gemma4
+- sft
+- finetune
+- reasoning
+- roleplay
+- tool-use
 ---
+<!doctype html>
+<html lang="en">
+    <head>
+        <meta charset="UTF-8" />
+        <meta name="viewport" content="width=device-width, initial-scale=1.0" />
+        <title>SpoomplesMaxx Whiskeyjack 12B</title>
+    </head>
+    <style>
+        @import url("https://fonts.googleapis.com/css2?family=Consolas&display=swap");
+        .crt-container {
+            padding: 10px;
+            max-width: 1000px;
+            margin: 0 auto;
+            width: 95%;
+        }
+        .crt-case {
+            background: #e8d7c3;
+            border-radius: 10px;
+            padding: 15px;
+            box-shadow:
+                inset -2px -2px 5px rgba(0, 0, 0, 0.3),
+                2px 2px 5px rgba(0, 0, 0, 0.2);
+        }
+        .crt-inner-case {
+            background: #e8d7c3;
+            border-radius: 8px;
+            padding: 3px;
+            box-shadow:
+                inset -1px -1px 4px rgba(0, 0, 0, 0.3),
+                1px 1px 4px rgba(0, 0, 0, 0.2);
+        }
+        .crt-bezel {
+            background: linear-gradient(145deg, #1a1a1a, #2a2a2a);
+            padding: 15px;
+            border-radius: 5px;
+            border: 3px solid #0a0a0a;
+            position: relative;
+            box-shadow:
+                inset 0 0 20px rgba(0, 0, 0, 0.5),
+                inset 0 0 4px rgba(0, 0, 0, 0.4),
+                inset 2px 2px 4px rgba(255, 255, 255, 0.05),
+                inset -2px -2px 4px rgba(0, 0, 0, 0.8),
+                0 0 2px rgba(0, 0, 0, 0.6),
+                -1px -1px 4px rgba(255, 255, 255, 0.1),
+                1px 1px 4px rgba(0, 0, 0, 0.3);
+        }
+        .crt-bezel::before {
+            content: "";
+            position: absolute;
+            top: 0;
+            left: 0;
+            right: 0;
+            bottom: 0;
+            background: linear-gradient(
+                45deg,
+                rgba(255, 255, 255, 0.03) 0%,
+                rgba(255, 255, 255, 0) 40%,
+                rgba(0, 0, 0, 0.1) 60%,
+                rgba(0, 0, 0, 0.2) 100%
+            );
+            border-radius: 3px;
+            pointer-events: none;
+        }
+        .terminal-screen {
+            background: #0c100d;
+            padding: 20px;
+            border-radius: 15px;
+            position: relative;
+            overflow: hidden;
+            font-family: "Consolas", monospace;
+            font-size: clamp(12px, 1.5vw, 16px);
+            color: #3dc862;
+            line-height: 1.4;
+            text-shadow: 0 0 2px #3dc862;
+            filter: brightness(1.1) contrast(1.1);
+            box-shadow:
+                inset 0 0 30px rgba(0, 0, 0, 0.9),
+                inset 0 0 8px rgba(0, 0, 0, 0.8),
+                0 0 5px rgba(0, 0, 0, 0.6);
+            max-width: 80ch;
+            margin: 0 auto;
+        }
+        .terminal-screen h2,
+        .terminal-screen h3 {
+            font-size: clamp(16px, 2vw, 20px);
+            margin-bottom: 1em;
+            color: #ffdf00;
+            text-shadow: 0 0 3px rgba(255, 223, 0, 0.5);
+        }
+        .terminal-screen pre.code-block-image {
+            display: inline-block;
+            text-align: left;
+            font-size: clamp(2px, 0.4vw, 12px);
+            font-family: monospace;
+            margin: 1em 0;
+            background-color: #1a1a1a;
+            padding: 1em;
+            border-radius: 4px;
+            color: #3dc862;
+            overflow-x: auto;
+            line-height: 1;
+            max-width: 100%;
+            overflow: hidden;
+            white-space: pre;
+        }
+        .terminal-screen pre.code-block {
+            display: inline-block;
+            text-align: left;
+            font-size: clamp(10px, 1.3vw, 14px);
+            font-family: monospace;
+            margin: 1em 0;
+            background-color: #1a1a1a;
+            padding: 1em;
+            border-radius: 4px;
+            color: #3dc862;
+            overflow-x: auto;
+            line-height: 1;
+            max-width: 100%;
+            overflow: hidden;
+            white-space: pre;
+        }
+        .terminal-screen::before {
+            content: "";
+            position: absolute;
+            top: 0;
+            left: 0;
+            right: 0;
+            bottom: 0;
+            background:
+                linear-gradient(
+                    rgba(18, 16, 16, 0) 50%,
+                    rgba(0, 0, 0, 0.25) 50%
+                ),
+                url("data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAADIAAAAyBAMAAADsEZWCAAAAGFBMVEUAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA4o8JoAAAAB3RSTlMAGwQIEQMYADcPzwAAACJJREFUKM9jYBgFo2AU0Beg+A8YMCLxGYZCbNQEo4BaAAD5TQiR5wU9vAAAAABJRU5ErkJggg==");
+            background-size: 100% 2.5px;
+            pointer-events: none;
+            z-index: 2;
+        }
+        .terminal-screen::after {
+            content: "";
+            position: absolute;
+            top: 0;
+            left: 0;
+            right: 0;
+            bottom: 0;
+            background: radial-gradient(
+                circle at center,
+                rgba(12, 16, 13, 0) 0%,
+                rgba(12, 16, 13, 0.2) 50%,
+                rgba(12, 16, 13, 0.15) 100%
+            );
+            border-radius: 20px;
+            pointer-events: none;
+            z-index: 1;
+        }
+        .terminal-screen .notice {
+            margin: 1.5em 0;
+            padding: 0.8em 1.2em;
+            border: 1px solid #ffdf00;
+            border-radius: 4px;
+            background-color: rgba(255, 223, 0, 0.04);
+        }
+        .terminal-screen .notice h3 {
+            margin-top: 0.2em;
+            margin-bottom: 0.5em;
+        }
+        .terminal-screen .notice p {
+            margin-bottom: 0.2em;
+        }
+        .terminal-screen strong,
+        .terminal-screen em {
+            color: #f0f0f0;
+        }
+        .terminal-screen p,
+        .terminal-screen li {
+            color: #3dc862;
+        }
+        .terminal-screen a {
+            color: #5da9ff;
+            text-decoration: underline;
+            text-shadow: 0 0 2px rgba(93, 169, 255, 0.5);
+            transition: opacity 0.2s;
+        }
+        .terminal-screen a:hover {
+            opacity: 0.8;
+        }
+        .terminal-screen code,
+        .terminal-screen kbd,
+        .terminal-screen samp {
+            color: #3dc862;
+            font-family: "Consolas", monospace;
+            text-shadow: 0 0 2px #3dc862;
+            background-color: #1a1a1a;
+            padding: 0.2em 0.4em;
+            border-radius: 4px;
+        }
+    </style>
+    <div class="crt-container">
+        <div class="crt-case">
+            <div class="crt-inner-case">
+                <div class="crt-bezel">
+                    <div class="terminal-screen">
+                        <div style="text-align: center">
+                            <h2>SpoomplesMaxx-Whiskeyjack-12B</h2>
+                            <h3>"Camp Robber"</h3>
+                            <pre class="code-block-image">
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒░░░��▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░░░▓▓▓▓▓▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒░░░░░░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░▒▓▓▓▓▓▓░▓░░▒░▓▓▓▓▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░▓▓▓▓▓▓▓▓▓▓░░░▓▓▓▓░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▒░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░▒▒░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒▒▒▒▒▒▒▒▒▒░░░▓▓▓▓▓▓▓▓▓▓▓▓░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▓▓▓▓▓▓▓▓▓▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▓▓▓▓▓▓░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▓▓▓▓▓▓▓▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒░▓▓▓▓▓▓▓▓░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░▒▒░▒▒▒▒▒▒▒▒▒▒▒▒▒▒░░░▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░▒░▒░░▒░░░░░░▓▓▓▓▓▓▓▓▓▓░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░░░░▒░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░░░░▒░▒░░░░▓▓▓▓▓▓▓▓▓▓░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░▓░░░░░░░░░░░░▓░░▒░▓░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░░░░░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░▓░░░░░░▒░░░▒░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░▓░░░░░▓░░░░░▒░░░░░░▒▓▓▓▓▓▓▓▓▓▓▓▓░▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░░▓░░░░░▓░░░▒▓▓▓▓▓▓▓▓▓▓▓▓▒░▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░▓░░░▓░░░░░░░░░░░░░▒▒▓▓▓▓▓▓▓▓▓▓▓▓▒▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒░░░░░░░░░▒░░▒░░░▒▒▒▒▒▓▓▓▓▓▓▓▓▓▓░░▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓░▓░░░▒░░▓░▒▓░░░▒▒▒▒▒▓▓▓▓▓▓▓▓▓▓▓▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░▓▓▓▓▓▓▓▓▓▓▓░░▒░▓░░▓░░░░░▒▒▒▒▒▒▒▒▓▒▓▓▓▓▓▓░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░▓▓▓▓▓▓▓▓▓▓▒░░░░▓░▒░░░░▒▒▒▒▒▒▒▒▒▒▒▒▓░░▓░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░▓▓▓▓▓▒░░░░░░░▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░▓▓▓▓░░▓▓▓▓░░░▒░▓░░░░░░░░░░░░░░░▒░▓▓▓▓▓▓▓░░▓▓▓▓▓▓▓▓▓░░░░░░░░▒▓▓░░░░░░░░░░░▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░▓▓▓░░░▓▓▓░░░░▓░░░░░░░░▓▓▓▓▓▓▓▓░░▓▓▓▓▓▓▓▓▓░░░░▓░░░░░░░░▓▓▓▓▓░░▓▓▓▓░░░░░░░░░▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░▓▓░░░░▓▓░░░▓░░▒░░░░░▓▓▓▓▓▓▓▓▓▓▓░░▓▓▓▓▓▓░░░░░▓░░░░░▓▓▓▓▓▓▓▓▓▓░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░▓░░░░▓▓▓░▓░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓░▓▓▓░░░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓░░░░░░░░▓▓▓░░░▓▓▓▓▓▒░░░░░░░░▒▓▓▓▓▓▓▓▓▓▓▓▓▓░▓░▓░▓░░░░░▓░▓▓▓▓▓░░░▓▓▓▓▓▓▓▓▓░░░░░▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓░░░░░░░░░▓▓░░▓▓▓▓▓▒░░░░░░▓░░░▓▓▓▓▓▓▓▓░░░░░░░▓░���░▓▓▓▓▓▓▓▓▓▓▓▓▓░▓▒░░░▒▓▓▓▓▓░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓░░░░░░░▓░▓▓▓▓▓▒░░░░░░░░░░▓▓▓▓▓░░░░░░░░░░▓░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░▓▓░░░░░░▓▓▓▒░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░▓▓▓▓▓░░░░░░░░░▓▓▓░░░░░░░░░░▓▓▓▓░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░▓▓░░░░░░▓▓▓▓▓░░▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░▓▓▓▓░░░░░░░░░▓░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░▓▓░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░▓▓▓▓░░░░░░░░░░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░▓▓▓░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░▓▓▓░░░░░░░▓░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░▓▓▓▓▓░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░▓▓░░░░░░░░▒░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░▓▓▓▓▓▓▓▓▓░▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░▒░░▒░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░▓▓▓▓▓▓▓▓▓▓▓▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░░░░░░░░▓░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓░░░░░░░▓░░░▓░░░░░░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓░░░░░░░░▓░░░▒░░░░░░▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓░░░░░░▓▓▓░░▒░░░▒░░▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓░░▓▓▓▓▓▓░▒░░░░░▒▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░░▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓��▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
+                            </pre>
+                        </div>
+                        <p>
+                            SpoomplesMaxx is a generalist model line with
+                            primary strengths in creative writing and roleplay,
+                            plus competence at instruction following,
+                            reasoning, and tool calling. Whiskeyjack brings the
+                            corvid line to Gemma: a full-parameter SFT of
+                            <strong>gemma-4-12B</strong>, trained in both
+                            thinking and non-thinking modes, with thinking off
+                            by default.
+                        </p>
+                        <p>
+                            Named for <em>Perisoreus canadensis</em> — the
+                            Canada jay, better known as the whisky jack or camp
+                            robber. A corvid bold enough to land on your hand
+                            and fly off with your lunch. The 35B got the
+                            jackdaw; the 12B gets the smaller, friendlier
+                            thief.
+                        </p>
+                        <h3>Prompt format</h3>
+                        <p>
+                            Gemma 4 uses a new turn format. It shares nothing
+                            with Gemma 3 — there is no
+                            <code>&lt;start_of_turn&gt;</code> — and the
+                            assistant role is spelled <code>model</code>:
+                        </p>
+                        <pre class="code-block">
+&lt;|turn&gt;user
+...&lt;turn|&gt;
+&lt;|turn&gt;model
+&lt;|channel&gt;thought
 ...reasoning...
+&lt;channel|&gt;...answer...&lt;turn|&gt;
+</pre>
+                        <pre class="code-block">
+STOPS:   stop on &lt;turn|&gt; (id 106). &lt;eos&gt; (id 1) is kept as a secondary
+         EOS, but never set &lt;eos&gt; alone -- turns end on &lt;turn|&gt;.
+</pre>
+                        <p>
+                            The control tokens (<code>&lt;turn|&gt;</code>,
+                            <code>&lt;|channel&gt;</code>/<code>&lt;channel|&gt;</code>,
+                            <code>&lt;|tool_call&gt;</code>/<code>&lt;tool_call|&gt;</code>)
+                            were audited before training and re-verified after
+                            it: stop battery, boundary probes, and a tool-call
+                            battery all pass on the published checkpoint.
+                        </p>
+                        <h3>Thinking behavior</h3>
+                        <p>
+                            Thinking is opt-in and <strong>off by
+                            default</strong>. A <code>&lt;|think|&gt;</code>
+                            marker at the top of the <strong>system</strong>
+                            turn switches it on;
+                            <code>apply_chat_template(enable_thinking=True)</code>
+                            injects it for you. Both modes share the same bare
+                            <code>&lt;|turn&gt;model</code> generation prefix —
+                            the model decides on its own whether to open
+                            <code>&lt;|channel&gt;thought</code>.
+                        </p>
+                        <pre class="code-block">
+MODE CONTROL:
+  (default)               thinking OFF -- no marker, no thought channel
+  enable_thinking=True    injects &lt;|think|&gt; into the system turn; the
+                          model opens &lt;|channel&gt;thought on its own
+PARSER NOTE:   reasoning sits between &lt;|channel&gt;thought and &lt;channel|&gt;;
+               the visible answer follows &lt;channel|&gt; in the same turn
+</pre>
+                        <div class="notice">
+                            <h3>The chat template is not stock Gemma 4</h3>
+                            <p>
+                                Upstream Gemma 4 appends an empty thought
+                                channel
+                                (<code>&lt;|channel&gt;thought\n&lt;channel|&gt;</code>)
+                                to non-thinking turns. That form shows up
+                                <strong>0 times in 1,000 training turns</strong>
+                                of this corpus — a no-thoughts turn simply
+                                carries no channel — so the template here drops
+                                it. The stock template ships alongside as
+                                <code>chat_template.gemma-it-original.jinja</code>.
+                                Restore it and you push the model out of
+                                distribution: reasoning leaks into the answer
+                                and tool calls lose their opener.
+                            </p>
+                        </div>
+                        <p>
+                            What the thoughts look like depends on the system
+                            prompt. Under a SillyTavern-style character card
+                            the model writes a structured planner (~750 chars;
+                            23/23 of the cards that opened a channel). Under
+                            the corpus's own RP framing it writes short
+                            first-person interiority (~90 chars). The model
+                            learned both forms separately, and the prompt picks
+                            which one you get.
+                        </p>
+                        <p>The planner, when it shows up:</p>
+                        <pre class="code-block">
+SCENE: where/when, atmosphere, key environmental details currently in play
+CHARACTERS: who is present and their current physical/emotional state and motivation
+CONTINUITY: established facts that must stay consistent
+THREADS: active tensions and where they stand right now
+PLAN: what THIS turn needs to accomplish and the approach it takes
+</pre>
+                        <p>
+                            One more thing to expect: a conversational
+                            companion persona usually produces no thought
+                            channel at all (0/6 in testing), even with thinking
+                            on. Companion rows in the corpus are mostly
+                            non-thinking, and the model follows the data.
+                        </p>
+                        <h3>Tool calling</h3>
+                        <p>
+                            Gemma 4 tool calls use a DSL, <strong>not
+                            JSON</strong>:
+                        </p>
+                        <pre class="code-block">
+FORM:     &lt;|tool_call&gt;call:NAME{key:&lt;|"|&gt;value&lt;|"|&gt;}&lt;tool_call|&gt;
+EXAMPLE:  &lt;|tool_call&gt;call:get_weather{city:&lt;|"|&gt;Lisbon&lt;|"|&gt;}&lt;tool_call|&gt;
+</pre>
+                        <div class="notice">
+                            <h3>Serve tool calls inside one turn</h3>
+                            <p>
+                                In the training corpus a whole tool episode
+                                lives inside a single
+                                <code>&lt;|turn&gt;model</code>, with
+                                <code>&lt;|tool_response&gt;</code> blocks
+                                interleaved inline. The model never emitted
+                                <code>&lt;turn|&gt;</code> after a call, so it
+                                never learned to yield there. A harness that
+                                waits for <code>&lt;turn|&gt;</code> will hang
+                                while the model keeps generating plausible
+                                calls — the classic infinite tool loop.
+                            </p>
+                            <pre class="code-block">
+SERVE WITH:  stop=["&lt;tool_call|&gt;"]
+THEN:        inject &lt;|tool_response&gt;response:NAME{...}&lt;tool_response|&gt;
+             and continue the SAME turn
+NEVER:       wait for &lt;turn|&gt; after a tool call
+</pre>
+                        </div>
+                        <h3>Key Details</h3>
+                        <pre class="code-block">
+BASE MODEL: google/gemma-4-12B
+LICENSE:    gemma
+NOTE:       the base is multimodal, so the checkpoint loads with
+            AutoModelForImageTextToText (see Quickstart)</pre>
+                        <h3>Training</h3>
+                        <pre class="code-block">
+METHOD:   FULL-PARAMETER SFT -- ms-swift (swift sft), DeepSpeed ZeRO-2,
+          torch SDPA attention, custom liger fused CE
+STAGES:   three, each tagged in this repo; main = stage 3
+stage 1 (v1-baseline-rp)   aviary burn corpus, 1 epoch
+                           1,917 steps @ lr 1e-5  eval 1.311   tok-acc 0.6465
+stage 2 (v2-corrected-rp)  thinking-weighted resample
+                           568 steps @ lr 2e-6    eval 1.3067  tok-acc 0.6479
+stage 3 (main)             + 4,000 converted RP-reasoning rows
+                           574 steps @ lr 2e-6    eval 1.301   tok-acc 0.6491
+</pre>
+                        <div class="notice">
+                            <h3>Why there is a stage 3</h3>
+                            <p>
+                                Stage 2 could think, but only under one prompt
+                                shape: 19,605 of the corpus's 20,666
+                                thought-bearing rows share a single RP framing.
+                                So the model opened a thought channel on 8/8
+                                in-corpus rows — and on 1 of 25 real character
+                                cards. Stage 3 mixed in RP-reasoning rows under
+                                ~4,000 distinct character cards so that
+                                thinking no longer depends on one specific
+                                prompt.
+                            </p>
+                            <pre class="code-block">
+opens a thought channel on 25 held-out character cards
+(846-5,053 chars, short opening message):
+  stage 2:   1/25   (4%)
+  stage 3:  23/25  (92%)
+unchanged across the pass:
+  stop rate 10/10 in both thinking and non-thinking modes
+  tool-call round trip passes
+  stray channels with thinking off: 0/3
+  P(&lt;channel|&gt;) at the true close: 1.000
+</pre>
+                        </div>
+                        <h3>Sampling</h3>
+                        <p>
+                            Use the defaults in <code>generation_config.json</code>.
+                            <pre class="code-block">
+"temperature": 1.0,
+"top_k": 64,
+"top_p": 0.95,
+                            </pre>
+                        </p>
+                        <h3>Quickstart</h3>
+                        <pre class="code-block">
 from transformers import AutoModelForImageTextToText, AutoTokenizer
 tok = AutoTokenizer.from_pretrained("aimeri/spoomplesmaxx-whiskeyjack-12B")
+model = AutoModelForImageTextToText.from_pretrained(
+    "aimeri/spoomplesmaxx-whiskeyjack-12B",
+    dtype="bfloat16", device_map="auto")
 msgs = [{"role": "user", "content": "Solve (x + 2)^2 = 0."}]
+ids = tok.apply_chat_template(msgs, add_generation_prompt=True,
+    enable_thinking=True, return_tensors="pt").to(model.device)
 out = model.generate(ids, max_new_tokens=512)
 print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=False))
+                        </pre>
+                    </div>
+                </div>
+            </div>
+        </div>
+    </div>
+</html>