199 lines
6.4 KiB
JSON
199 lines
6.4 KiB
JSON
{
|
|
"version": 1,
|
|
"rules": [
|
|
{
|
|
"id": "qwen36_hauhau_multimodal",
|
|
"priority": 100,
|
|
"match_any": [
|
|
"Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive"
|
|
],
|
|
"family_label": "Qwen3.6 35B A3B HauhauCS aggressive",
|
|
"sources": [
|
|
"HauhauCS model card",
|
|
"Qwen3.5 official model card"
|
|
],
|
|
"support_notes": {
|
|
"hf": "Not in the HF lineup here. Use the GGUF nodes for this model.",
|
|
"gguf": "Hard toggle via enable_thinking. Pair the vision node with the matching mmproj file.",
|
|
"soft_switch": "No official /think or /no_think soft switch for this family."
|
|
},
|
|
"thinking_on": {
|
|
"temperature": 1.0,
|
|
"top_p": 0.95,
|
|
"top_k": 20,
|
|
"max_tokens": 32768,
|
|
"ctx_min": 131072
|
|
},
|
|
"thinking_off": {
|
|
"temperature": 0.7,
|
|
"top_p": 0.8,
|
|
"top_k": 20,
|
|
"max_tokens": 32768,
|
|
"ctx_min": 131072
|
|
},
|
|
"notes": [
|
|
"Provider says keep at least 128K context to preserve thinking quality.",
|
|
"Q4_K_P is the best quality-size starting point; move to Q6_K_P or Q8_K_P when VRAM allows.",
|
|
"IQ quants are fallback options when you cannot fit Q4_K_M or Q4_K_P."
|
|
]
|
|
},
|
|
{
|
|
"id": "qwen35_family",
|
|
"priority": 80,
|
|
"match_any": [
|
|
"Qwen3.5",
|
|
"qwen3.5"
|
|
],
|
|
"family_label": "Qwen3.5 family",
|
|
"sources": [
|
|
"Qwen3.5 official model card"
|
|
],
|
|
"support_notes": {
|
|
"hf": "Hard toggle via enable_thinking when the backend honors the chat template.",
|
|
"gguf": "Hard toggle via enable_thinking in the Qwen chat template.",
|
|
"soft_switch": "No official /think or /no_think soft switch for Qwen3.5."
|
|
},
|
|
"thinking_on": {
|
|
"temperature": 1.0,
|
|
"top_p": 0.95,
|
|
"top_k": 20,
|
|
"max_tokens": 32768,
|
|
"ctx_min": 131072
|
|
},
|
|
"thinking_off": {
|
|
"temperature": 0.7,
|
|
"top_p": 0.8,
|
|
"top_k": 20,
|
|
"max_tokens": 32768,
|
|
"ctx_min": 131072
|
|
},
|
|
"notes": [
|
|
"For precise coding, Qwen recommends temperature 0.6 with thinking still enabled.",
|
|
"For hard non-thinking reasoning tasks, Qwen recommends top_p 1.0 and top_k 40 when the backend exposes them.",
|
|
"Provider says keep at least 128K context to preserve thinking capability and use 32K+ output budget when possible."
|
|
]
|
|
},
|
|
{
|
|
"id": "qwen3_family",
|
|
"priority": 70,
|
|
"match_any": [
|
|
"Qwen3VL-",
|
|
"qwen3vl-",
|
|
"Qwen3-VL",
|
|
"Qwen3-",
|
|
"qwen3-"
|
|
],
|
|
"exclude_any": [
|
|
"Qwen3.5",
|
|
"qwen3.5",
|
|
"Qwen3.6",
|
|
"qwen3.6"
|
|
],
|
|
"family_label": "Qwen3 family",
|
|
"sources": [
|
|
"Qwen3 official model card",
|
|
"Qwen Transformers docs"
|
|
],
|
|
"support_notes": {
|
|
"hf": "Hard toggle via enable_thinking, with /think and /no_think available as soft switches.",
|
|
"gguf": "Hard toggle via enable_thinking. /think and /no_think are available only on Qwen3 chat-template builds.",
|
|
"soft_switch": "Qwen3 officially supports /think and /no_think when thinking is enabled."
|
|
},
|
|
"thinking_on": {
|
|
"temperature": 0.6,
|
|
"top_p": 0.95,
|
|
"top_k": 20,
|
|
"max_tokens": 32768,
|
|
"ctx_min": 32768
|
|
},
|
|
"thinking_off": {
|
|
"temperature": 0.7,
|
|
"top_p": 0.8,
|
|
"top_k": 20,
|
|
"max_tokens": 32768,
|
|
"ctx_min": 32768
|
|
},
|
|
"notes": [
|
|
"Qwen warns against greedy decoding for thinking mode because it can repeat or degrade quality.",
|
|
"Provider recommends 32K output budget for most tasks and only enabling YaRN when you really need longer than the native 32K context.",
|
|
"For harder reasoning benchmarks, Qwen suggests stretching output budget beyond 32K if your backend allows it."
|
|
]
|
|
},
|
|
{
|
|
"id": "gemma4_large_family",
|
|
"priority": 60,
|
|
"match_any": [
|
|
"Gemma-4-26B",
|
|
"gemma-4-26B",
|
|
"Gemma-4-31B",
|
|
"gemma-4-31B"
|
|
],
|
|
"family_label": "Gemma 4 large family",
|
|
"sources": [
|
|
"Gemma 4 official model card"
|
|
],
|
|
"support_notes": {
|
|
"hf": "Gemma 4 supports configurable thinking, but current repo behavior can vary by backend.",
|
|
"gguf": "Use the settings manually and verify behavior in RAW_TRACE because GGUF support varies by backend.",
|
|
"soft_switch": "Gemma uses chat-template thinking controls instead of /think or /no_think."
|
|
},
|
|
"thinking_on": {
|
|
"temperature": 1.0,
|
|
"top_p": 0.95,
|
|
"top_k": 64,
|
|
"max_tokens": 1024,
|
|
"ctx_min": 262144
|
|
},
|
|
"thinking_off": {
|
|
"temperature": 1.0,
|
|
"top_p": 0.95,
|
|
"top_k": 64,
|
|
"max_tokens": 1024,
|
|
"ctx_min": 262144
|
|
},
|
|
"notes": [
|
|
"Gemma 4 uses one standardized sampler for both thinking and non-thinking modes.",
|
|
"Official context target is 256K on 26B A4B and 31B variants.",
|
|
"If the backend does not expose top_k, keep temperature and top_p at the official defaults."
|
|
]
|
|
},
|
|
{
|
|
"id": "gemma4_small_family",
|
|
"priority": 50,
|
|
"match_any": [
|
|
"Gemma-4-E2B",
|
|
"gemma-4-E2B",
|
|
"Gemma-4-E4B",
|
|
"gemma-4-E4B"
|
|
],
|
|
"family_label": "Gemma 4 small family",
|
|
"sources": [
|
|
"Gemma 4 official model card"
|
|
],
|
|
"support_notes": {
|
|
"hf": "Gemma 4 supports configurable thinking, but current repo behavior can vary by backend.",
|
|
"gguf": "Use the settings manually and verify behavior in RAW_TRACE because GGUF support varies by backend.",
|
|
"soft_switch": "Gemma uses chat-template thinking controls instead of /think or /no_think."
|
|
},
|
|
"thinking_on": {
|
|
"temperature": 1.0,
|
|
"top_p": 0.95,
|
|
"top_k": 64,
|
|
"max_tokens": 1024,
|
|
"ctx_min": 131072
|
|
},
|
|
"thinking_off": {
|
|
"temperature": 1.0,
|
|
"top_p": 0.95,
|
|
"top_k": 64,
|
|
"max_tokens": 1024,
|
|
"ctx_min": 131072
|
|
},
|
|
"notes": [
|
|
"Gemma 4 uses one standardized sampler for both thinking and non-thinking modes.",
|
|
"Official context target is 128K on E2B and E4B variants.",
|
|
"Gemma recommends higher visual token budgets only for OCR or fine detail tasks; lower budgets are faster for captioning or video summaries."
|
|
]
|
|
}
|
|
]
|
|
} |