Initial SQL-only 1C adapter baseline

2026-07-22 03:03:47 +03:00
commit e2503b77e7
545 changed files with 184711 additions and 0 deletions
@@ -0,0 +1 @@
+
@@ -0,0 +1,29 @@
+id: animagine-xl-4_0
+name: Animagine XL 4.0
+type: image-diffusion-model
+task:
+  - image-generation
+language:
+  - en
+  - multilingual
+source: huggingface
+upstream_id: cagliostrolab/animagine-xl-4.0
+upstream_url: https://huggingface.co/cagliostrolab/animagine-xl-4.0
+license: openrail++
+status: candidate
+created_at: 2026-07-04
+storage_path: /models/image/animagine-xl-4.0
+format: diffusers
+quantization: fp16
+context_length: null
+vram_required_gb: 12
+parameters: null
+base_model: stabilityai/stable-diffusion-xl-base-1.0
+adapter_for: null
+datasets: []
+eval_suites: []
+deployment:
+  target: docker-gpu.cin.su
+  runtime: transformers
+  served_model_name: animagine-xl-4
+notes: "Anime-focused SDXL finetune with Diffusers support on Hugging Face. Good fit for RTX 4090 and compatible with the current StableDiffusionXLPipeline-based image service."
@@ -0,0 +1,38 @@
+id: devstral-small-2-24b-instruct-2512-q4_k_m
+name: Devstral Small 2 24B Instruct 2512 GGUF Q4_K_M
+type: gguf-model
+task:
+  - 1c
+  - code
+  - agentic-coding
+  - text
+  - tool-use
+language:
+  - ru
+  - en
+source: huggingface
+upstream_id: bartowski/mistralai_Devstral-Small-2-24B-Instruct-2512-GGUF
+upstream_url: https://huggingface.co/bartowski/mistralai_Devstral-Small-2-24B-Instruct-2512-GGUF
+license: apache-2.0
+status: candidate
+created_at: 2026-06-18
+storage_path: /models/gguf/1c/devstral-small-2-24b-instruct-2512-q4_k_m
+filename: mistralai_Devstral-Small-2-24B-Instruct-2512-Q4_K_M.gguf
+file_size_bytes: 14334438272
+format: gguf
+quantization: Q4_K_M
+context_length: null
+deployment_context_length: 32768
+vram_required_gb: null
+parameters: 24B
+base_model: mistralai/Devstral-Small-2-24B-Instruct-2512
+adapter_for: null
+datasets: []
+eval_suites:
+  - plugins/1c/evals/smoke.yaml
+deployment:
+  target: docker-gpu.cin.su
+  runtime: llama.cpp
+  served_model_name: devstral-1c-q4
+  compose: core/deploy/docker-gpu/llama-cpp/compose.yaml
+notes: "GGUF Q4_K_M quant for 1C/code experiments. Loads on docker-gpu.cin.su with llama.cpp CUDA and uses GPU, but current llama.cpp server returns empty decoded content for chat/completion despite token generation. Runtime is experimental until a compatible template/runtime/quant is found."
@@ -0,0 +1,29 @@
+id: illustriousxl
+name: Illustrious XL
+type: image-diffusion-model
+task:
+  - image-generation
+language:
+  - en
+  - multilingual
+source: huggingface
+upstream_id: glides/illustriousxl
+upstream_url: https://huggingface.co/glides/illustriousxl
+license: mit
+status: candidate
+created_at: 2026-07-04
+storage_path: /models/image/illustriousxl
+format: diffusers
+quantization: fp16
+context_length: null
+vram_required_gb: 12
+parameters: null
+base_model: stabilityai/stable-diffusion-xl-base-1.0
+adapter_for: null
+datasets: []
+eval_suites: []
+deployment:
+  target: docker-gpu.cin.su
+  runtime: transformers
+  served_model_name: illustriousxl
+notes: "Stylized illustration / anime-oriented SDXL model with explicit StableDiffusionXLPipeline usage on Hugging Face. Useful as a third visual style alongside photorealism and pure anime."
@@ -0,0 +1,30 @@
+id: lmt-60-4b
+name: LMT-60 4B
+type: translation-model
+task:
+  - translation
+language:
+  - ru
+  - en
+  - multilingual
+source: huggingface
+upstream_id: NiuTrans/LMT-60-4B
+upstream_url: https://huggingface.co/NiuTrans/LMT-60-4B
+license: apache-2.0
+status: candidate
+created_at: 2026-06-19
+storage_path: /models/translation/lmt-60-4b
+format: safetensors
+quantization: none
+context_length: null
+vram_required_gb: null
+parameters: 4B
+base_model: null
+adapter_for: null
+datasets: []
+eval_suites: []
+deployment:
+  target: docker-gpu.cin.su
+  runtime: transformers
+  served_model_name: lmt-60-4b
+notes: "Apache-2.0 multilingual translation candidate. Chosen over NLLB for fewer license restrictions while staying below larger 8B translation models."
@@ -0,0 +1,32 @@
+id: qwen-image-edit
+name: Qwen Image Edit
+type: image-diffusion-model
+task:
+  - image-editing
+  - inpainting
+  - image-generation
+language:
+  - ru
+  - en
+  - multilingual
+source: huggingface
+upstream_id: Qwen/Qwen-Image-Edit
+upstream_url: https://huggingface.co/Qwen/Qwen-Image-Edit
+license: apache-2.0
+status: candidate
+created_at: 2026-06-19
+storage_path: /models/image/qwen-image-edit
+format: diffusers
+quantization: bf16
+context_length: null
+vram_required_gb: 24
+parameters: 20B
+base_model: Qwen/Qwen-Image
+adapter_for: null
+datasets: []
+eval_suites: []
+deployment:
+  target: docker-gpu.cin.su
+  runtime: transformers
+  served_model_name: qwen-image-edit
+notes: "Heavy image editing candidate with stronger text/rendering/editing capabilities than SDXL. On 2026-06-20 it loaded as QwenImageEditPipeline in about 31s on RTX 4090 with CPU offload, but a 512x512 1-step edit did not finish within 1800s. Keep it as an experimental/manual-switch model until a quantized runtime or larger GPU is available."
@@ -0,0 +1,33 @@
+id: qwen2_5-vl-7b-instruct
+name: Qwen2.5-VL 7B Instruct
+type: vision-language-model
+task:
+  - video
+  - image-understanding
+  - document-understanding
+  - visual-question-answering
+language:
+  - ru
+  - en
+  - multilingual
+source: huggingface
+upstream_id: Qwen/Qwen2.5-VL-7B-Instruct
+upstream_url: https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct
+license: apache-2.0
+status: candidate
+created_at: 2026-06-19
+storage_path: /models/video/qwen2.5-vl-7b-instruct
+format: safetensors
+quantization: none
+context_length: null
+vram_required_gb: null
+parameters: 7B
+base_model: Qwen/Qwen2.5-VL
+adapter_for: null
+datasets: []
+eval_suites: []
+deployment:
+  target: docker-gpu.cin.su
+  runtime: transformers
+  served_model_name: qwen2.5-vl-7b-instruct
+notes: "Video/image understanding candidate for the video plugin. Supports long-video understanding according to the model card."
@@ -0,0 +1,38 @@
+id: qwen3-14b-instruct-q6_k
+name: Qwen3 14B Instruct GGUF Q6_K
+type: gguf-model
+task:
+  - text
+  - chat
+  - summarization
+  - code
+  - tool-use
+  - 1c-rag
+language:
+  - ru
+  - en
+  - multilingual
+source: huggingface
+upstream_id: Qwen/Qwen3-14B-GGUF
+upstream_url: https://huggingface.co/Qwen/Qwen3-14B-GGUF
+license: apache-2.0
+status: candidate
+created_at: 2026-06-19
+storage_path: /models/gguf/text/qwen3-14b-instruct-q6_k
+filename: Qwen3-14B-Q6_K.gguf
+file_size_bytes: 12121937248
+format: gguf
+quantization: Q6_K
+context_length: 40960
+deployment_context_length: 16384
+vram_required_gb: 16
+parameters: 14.8B
+base_model: Qwen/Qwen3-14B
+adapter_for: null
+datasets: []
+eval_suites: []
+deployment:
+  target: docker-gpu.cin.su
+  runtime: llama.cpp
+  served_model_name: qwen3-14b-q6
+notes: "Stronger general text/chat candidate than Qwen3 4B while still comfortable for a single RTX 4090 in GGUF Q6_K."
@@ -0,0 +1,32 @@
+id: qwen3-4b-1c-lora-v1
+name: Qwen3 4B 1C LoRA v1
+type: lora-adapter
+task:
+  - 1c
+  - bsl-code
+  - metadata-safety
+  - 1c-query
+  - explanation
+language:
+  - ru
+source: local-training
+license: internal
+status: draft
+created_at: 2026-06-18
+storage_path: /models/adapters/1c/qwen3-4b-1c-lora-v1
+format: safetensors
+quantization: null
+context_length: 32768
+vram_required_gb: null
+parameters: null
+base_model: qwen3-4b-instruct-2507
+adapter_for: qwen3-4b-instruct-2507
+datasets:
+  - plugins/1c/training/manifests/dataset.yaml
+eval_suites:
+  - plugins/1c/evals/smoke.yaml
+deployment:
+  target: docker-gpu.cin.su
+  runtime: vllm
+  served_model_name: qwen3-4b-1c
+notes: "Draft 1C LoRA adapter. Smoke training completed on docker-gpu.cin.su, then repeated with a 65-record synthetic/example dataset generated from 1C metadata and BSL snapshots; use it to validate the training/deployment pipeline, not as a quality 1C expert model until a real reviewed dataset is prepared."
@@ -0,0 +1,37 @@
+id: qwen3-4b-instruct-2507
+name: Qwen3 4B Instruct 2507
+type: base-model
+task:
+  - text
+  - chat
+  - summarization
+  - code
+  - tool-use
+  - 1c-rag
+language:
+  - ru
+  - en
+source: huggingface
+upstream_id: Qwen/Qwen3-4B-Instruct-2507
+upstream_url: https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507
+license: apache-2.0
+status: staging
+created_at: 2026-06-18
+storage_path: /models/base/qwen3-4b-instruct-2507
+format: safetensors
+quantization: none
+context_length: 262144
+deployment_context_length: 32768
+vram_required_gb: null
+parameters: 4B
+base_model: null
+adapter_for: null
+datasets: []
+eval_suites:
+  - evals/text/smoke.yaml
+deployment:
+  target: docker-gpu.cin.su
+  runtime: vllm
+  served_model_name: qwen3-4b-instruct
+  compose: core/deploy/docker-gpu/vllm/compose.yaml
+notes: "First base text model for local inference. Start with 32K context to reduce OOM risk; increase after GPU VRAM is confirmed."
@@ -0,0 +1,34 @@
+id: qwen3-coder-30b-a3b-1c-lora-v1
+name: Qwen3 Coder 30B A3B 1C LoRA v1
+type: lora-adapter
+task:
+  - 1c
+  - bsl-code
+  - metadata-safety
+  - metadata-write-verified
+  - metadata-write-learning-plan
+  - 1c-query
+  - explanation
+language:
+  - ru
+source: local-training
+license: internal
+status: draft
+created_at: 2026-07-04
+storage_path: /models/adapters/1c/qwen3-coder-30b-a3b-1c-lora-v1
+format: safetensors
+quantization: null
+context_length: 262144
+vram_required_gb: null
+parameters: null
+base_model: qwen3-coder-30b-a3b-instruct
+adapter_for: qwen3-coder-30b-a3b-instruct
+datasets:
+  - plugins/1c/training/manifests/dataset.yaml
+eval_suites:
+  - plugins/1c/evals/smoke.yaml
+deployment:
+  target: docker-gpu.cin.su
+  runtime: llama.cpp
+  served_model_name: qwen3-coder-1c-q6
+notes: "Draft adapter for the current Qwen3-Coder Q6 route. Train against the HF base, then convert or merge for the llama.cpp GGUF deployment used on docker-gpu.cin.su."
@@ -0,0 +1,38 @@
+id: qwen3-coder-30b-a3b-instruct-q4_k_m
+name: Qwen3 Coder 30B A3B Instruct GGUF Q4_K_M
+type: gguf-model
+task:
+  - 1c
+  - code
+  - agentic-coding
+  - repository-analysis
+  - tool-use
+language:
+  - ru
+  - en
+source: huggingface
+upstream_id: lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-GGUF
+upstream_url: https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-GGUF
+license: apache-2.0
+status: candidate
+created_at: 2026-06-19
+storage_path: /models/gguf/1c/qwen3-coder-30b-a3b-instruct-q4_k_m
+filename: Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf
+file_size_bytes: 18632186176
+format: gguf
+quantization: Q4_K_M
+context_length: 262144
+deployment_context_length: 32768
+vram_required_gb: null
+parameters: 30.5B total / 3.3B active
+base_model: Qwen/Qwen3-Coder-30B-A3B-Instruct
+adapter_for: null
+datasets: []
+eval_suites:
+  - plugins/1c/evals/smoke.yaml
+deployment:
+  target: docker-gpu.cin.su
+  runtime: llama.cpp
+  served_model_name: qwen3-coder-1c-q4
+  compose: core/deploy/docker-gpu/llama-cpp/compose.yaml
+notes: "Best current candidate found for 1C/code experiments. Not 1C-specific, but strong for agentic coding, long context, repository-scale understanding, and tool workflows."
@@ -0,0 +1,38 @@
+id: qwen3-coder-30b-a3b-instruct-q6_k
+name: Qwen3 Coder 30B A3B Instruct GGUF Q6_K
+type: gguf-model
+task:
+  - 1c
+  - code
+  - agentic-coding
+  - repository-analysis
+  - tool-use
+language:
+  - ru
+  - en
+source: huggingface
+upstream_id: lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-GGUF
+upstream_url: https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-GGUF
+license: apache-2.0
+status: candidate
+created_at: 2026-06-19
+storage_path: /models/gguf/1c/qwen3-coder-30b-a3b-instruct-q6_k
+filename: Qwen3-Coder-30B-A3B-Instruct-Q6_K.gguf
+file_size_bytes: 25104724288
+format: gguf
+quantization: Q6_K
+context_length: 262144
+deployment_context_length: 8192
+vram_required_gb: null
+parameters: 30.5B total / 3.3B active
+base_model: Qwen/Qwen3-Coder-30B-A3B-Instruct
+adapter_for: null
+datasets: []
+eval_suites:
+  - plugins/1c/evals/smoke.yaml
+deployment:
+  target: docker-gpu.cin.su
+  runtime: llama.cpp
+  served_model_name: qwen3-coder-1c-q6
+  compose: core/deploy/docker-gpu/llama-cpp/compose.yaml
+notes: "High-quality Q6_K quant for local 1C/code experiments on a single RTX 4090. Expected to require partial CPU/RAM offload or a short context because the GGUF file is larger than 24 GiB VRAM. Test first at 8K context, then 16K if stable."
@@ -0,0 +1,35 @@
+id: qwen3-coder-30b-a3b-instruct
+name: Qwen3 Coder 30B A3B Instruct
+type: base-model
+task:
+  - 1c
+  - code
+  - agentic-coding
+  - repository-analysis
+  - tool-use
+language:
+  - ru
+  - en
+source: huggingface
+upstream_id: Qwen/Qwen3-Coder-30B-A3B-Instruct
+upstream_url: https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct
+license: apache-2.0
+status: staging
+created_at: 2026-07-04
+storage_path: /models/base/qwen3-coder-30b-a3b-instruct
+format: safetensors
+quantization: none
+context_length: 262144
+deployment_context_length: 8192
+vram_required_gb: null
+parameters: 30.5B total / 3.3B active
+base_model: null
+adapter_for: null
+datasets: []
+eval_suites:
+  - plugins/1c/evals/smoke.yaml
+deployment:
+  target: docker-gpu.cin.su
+  runtime: training
+  served_model_name: qwen3-coder-30b-a3b-instruct
+notes: "Training base for the current local 1C/code route. Use this HF checkpoint for LoRA fine-tuning, then convert or merge adapters for the Q6 GGUF deployment."
@@ -0,0 +1,33 @@
+id: qwen3-vl-8b-instruct
+name: Qwen3-VL 8B Instruct
+type: vision-language-model
+task:
+  - video
+  - image-understanding
+  - document-understanding
+  - visual-question-answering
+language:
+  - ru
+  - en
+  - multilingual
+source: huggingface
+upstream_id: Qwen/Qwen3-VL-8B-Instruct
+upstream_url: https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct
+license: apache-2.0
+status: candidate
+created_at: 2026-06-19
+storage_path: /models/video/qwen3-vl-8b-instruct
+format: safetensors
+quantization: none
+context_length: null
+vram_required_gb: 18
+parameters: 8B
+base_model: Qwen/Qwen3-VL
+adapter_for: null
+datasets: []
+eval_suites: []
+deployment:
+  target: docker-gpu.cin.su
+  runtime: transformers
+  served_model_name: qwen3-vl-8b-instruct
+notes: "Newer vision-language candidate for image, document, and video understanding tests. Should be tested separately from large text/image models."
@@ -0,0 +1,29 @@
+id: realvisxl-v5_0
+name: RealVisXL V5.0
+type: image-diffusion-model
+task:
+  - image-generation
+language:
+  - en
+  - multilingual
+source: huggingface
+upstream_id: SG161222/RealVisXL_V5.0
+upstream_url: https://huggingface.co/SG161222/RealVisXL_V5.0
+license: openrail++
+status: candidate
+created_at: 2026-07-04
+storage_path: /models/image/realvisxl-v5.0
+format: diffusers
+quantization: fp16
+context_length: null
+vram_required_gb: 12
+parameters: null
+base_model: stabilityai/stable-diffusion-xl-base-1.0
+adapter_for: null
+datasets: []
+eval_suites: []
+deployment:
+  target: docker-gpu.cin.su
+  runtime: transformers
+  served_model_name: realvisxl-v5
+notes: "Photorealistic SDXL finetune for local RTX 4090 image generation. The Hugging Face model card exposes Diffusers usage and targets photorealistic output."
@@ -0,0 +1,29 @@
+id: sdxl-base-1_0
+name: Stable Diffusion XL Base 1.0
+type: image-diffusion-model
+task:
+  - image-generation
+language:
+  - en
+  - multilingual
+source: huggingface
+upstream_id: stabilityai/stable-diffusion-xl-base-1.0
+upstream_url: https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
+license: openrail++
+status: staging
+created_at: 2026-06-19
+storage_path: /models/image/sdxl-base-1.0
+format: diffusers
+quantization: fp16
+context_length: null
+vram_required_gb: 12
+parameters: null
+base_model: null
+adapter_for: null
+datasets: []
+eval_suites: []
+deployment:
+  target: docker-gpu.cin.su
+  runtime: transformers
+  served_model_name: sdxl-image
+notes: "Diffusers SDXL text-to-image candidate for local photo generation on RTX 4090."
@@ -0,0 +1,30 @@
+id: sdxl-inpainting-1_0
+name: Stable Diffusion XL Inpainting 1.0
+type: image-diffusion-model
+task:
+  - image-editing
+  - inpainting
+language:
+  - en
+  - multilingual
+source: huggingface
+upstream_id: diffusers/stable-diffusion-xl-1.0-inpainting-0.1
+upstream_url: https://huggingface.co/diffusers/stable-diffusion-xl-1.0-inpainting-0.1
+license: openrail++
+status: staging
+created_at: 2026-06-19
+storage_path: /models/image/sdxl-inpainting-1.0
+format: diffusers
+quantization: fp16
+context_length: null
+vram_required_gb: 12
+parameters: null
+base_model: stabilityai/stable-diffusion-xl-base-1.0
+adapter_for: null
+datasets: []
+eval_suites: []
+deployment:
+  target: docker-gpu.cin.su
+  runtime: transformers
+  served_model_name: sdxl-image
+notes: "Diffusers SDXL inpainting candidate for local masked image editing."
@@ -0,0 +1,32 @@
+id: whisper-large-v3-turbo-russian
+name: Whisper Large v3 Turbo Russian
+type: speech-model
+task:
+  - speech-to-text
+  - speech-translation
+language:
+  - ru
+source: huggingface
+upstream_id: dvislobokov/whisper-large-v3-turbo-russian
+upstream_url: https://huggingface.co/dvislobokov/whisper-large-v3-turbo-russian
+license: mit
+status: candidate
+created_at: 2026-06-19
+storage_path: /models/audio/whisper-large-v3-turbo-russian
+format: safetensors
+quantization: none
+filename: model.safetensors
+file_size_bytes: 3235581408
+context_length: null
+vram_required_gb: 6
+parameters: null
+base_model: openai/whisper-large-v3-turbo
+adapter_for: null
+datasets:
+  - Mozilla Common Voice 17
+eval_suites: []
+deployment:
+  target: docker-gpu.cin.su
+  runtime: transformers
+  served_model_name: whisper-large-v3-turbo-russian
+notes: "Russian fine-tuned Whisper turbo candidate for comparing Russian ASR quality against the base multilingual Whisper turbo."
@@ -0,0 +1,33 @@
+id: whisper-large-v3-turbo
+name: Whisper Large v3 Turbo
+type: speech-model
+task:
+  - speech-to-text
+  - speech-translation
+language:
+  - ru
+  - en
+  - multilingual
+source: huggingface
+upstream_id: openai/whisper-large-v3-turbo
+upstream_url: https://huggingface.co/openai/whisper-large-v3-turbo
+license: mit
+status: candidate
+created_at: 2026-06-19
+storage_path: /models/audio/whisper-large-v3-turbo
+format: safetensors
+quantization: none
+filename: model.safetensors
+file_size_bytes: 1617824864
+context_length: null
+vram_required_gb: null
+parameters: null
+base_model: openai/whisper-large-v3
+adapter_for: null
+datasets: []
+eval_suites: []
+deployment:
+  target: docker-gpu.cin.su
+  runtime: transformers
+  served_model_name: whisper-large-v3-turbo
+notes: "ASR and speech translation model for the audio plugin."