mrleo1nid/swarm-assistent
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Ship Assistent 0.13.0: real QLoRA pipeline and GGUF Ollama register.

Browse Source 
Replace fake train loop with TRL SFTTrainer, HF column mapping with fiction preset,
safetensors to GGUF conversion, and ollama create using ollama_base plus ADAPTER.

Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Leonid Pershin
2026-08-22 14:45:07 +03:00
co-authored by Cursor
parent 1a03c3178f
commit 474a674a35
12 changed files with 1002 additions and 133 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Tabs/Text2Image/Assistent.html
+23 -7
View File
@@ -222,6 +222,17 @@
<div class="sa-hf-status" id="sa_hf_status" role="status"></div>
<div class="sa-hf-list" id="sa_hf_list"></div>
<div class="sa-hf-preview" id="sa_hf_preview" hidden></div>
<div class="sa-hf-mapping-row" id="sa_hf_mapping_row" hidden>
<span class="sa-settings-hint">Маппинг колонок:</span>
<label>Preset
<select id="sa_hf_mapping_preset" class="sa-select">
<option value="">— вручную —</option>
<option value="fiction_tags_text">Fiction: title/tags → text</option>
</select>
</label>
<label>User col <select id="sa_hf_user_col" class="sa-select"></select></label>
<label>Assistant col <select id="sa_hf_asst_col" class="sa-select"></select></label>
</div>
<div class="sa-hf-import-row" id="sa_hf_import_row" hidden>
<label>Лимит строк <input type="number" id="sa_hf_import_limit" min="1" max="5000" value="200" /></label>
<label>Соотношение внешних:своих <input type="number" id="sa_hf_mix_ratio" min="0" max="20" step="0.5" value="3" title="Сколько внешних примеров на один свой" /></label>
@@ -253,17 +264,22 @@
</div>
<div class="sa-train-form" id="sa_train_form_qlora" hidden>
<p class="sa-settings-hint">Base model — HF id (safetensors). Тренер скачает веса сам (нужен HF_TOKEN в User Settings).</p>
<label>HF base model <input type="text" id="sa_qlora_base" placeholder="meta-llama/Llama-3.2-3B-Instruct" /></label>
<label>Имя адаптера <input type="text" id="sa_qlora_name" placeholder="assistent-lora-v1" /></label>
<label>HF base model <input type="text" id="sa_qlora_base" placeholder="Qwen/Qwen2.5-7B-Instruct" /></label>
<label>Ollama base (FROM для ADAPTER)
<select id="sa_qlora_ollama_base" class="sa-select"><option value="">—</option></select>
</label>
<label>Имя модели в Ollama <input type="text" id="sa_qlora_name" placeholder="my-lora:v1" /></label>
<div class="sa-settings-row sa-knob-row">
<label>rank <input type="number" id="sa_qlora_rank" min="4" max="128" value="16" /></label>
<label>alpha <input type="number" id="sa_qlora_alpha" min="4" max="256" value="32" /></label>
<label>LR <input type="number" id="sa_qlora_lr" min="0.000001" max="0.01" step="0.00001" value="0.0002" /></label>
<label>epochs <input type="number" id="sa_qlora_epochs" min="1" max="20" value="3" /></label>
<label>seq_len <input type="number" id="sa_qlora_seq" min="512" max="8192" step="256" value="2048" /></label>
<label>max_samples <input type="number" id="sa_qlora_max_samples" min="0" max="50000" value="0" title="0 = все строки" /></label>
</div>
<label class="sa-check"><input type="checkbox" id="sa_qlora_4bit" checked /> 4-bit QLoRA</label>
<label>HF датасет (опционально) <input type="text" id="sa_qlora_hf_dataset" placeholder="owner/name — только для раннера" /></label>
<label>HF датасет (опционально, без локального датасета) <input type="text" id="sa_qlora_hf_dataset" placeholder="krplt/ru-fictext-nsfw" /></label>
<p class="sa-settings-hint">После обучения: safetensors → GGUF (convert script) → ollama create. Настрой пути в Настройки → Модели.</p>
<button type="button" class="basic-button sa-primary" id="sa_btn_qlora_start">Запустить QLoRA</button>
<button type="button" class="basic-button" id="sa_btn_qlora_cancel" hidden>Отменить</button>
</div>
@@ -323,15 +339,15 @@
<label>Python <input type="text" id="sa_runner_python" placeholder="python или полный путь" /></label>
<label>Тип тренера
<select id="sa_runner_kind" class="sa-select">
<option value="">— не настроен —</option>
<option value="llama-factory">LLaMA-Factory</option>
<option value="unsloth">Unsloth</option>
<option value="builtin" selected>Встроенный (train_qlora.py)</option>
<option value="custom">Custom command</option>
</select>
</label>
<label>GGUF base model (.gguf) <input type="text" id="sa_runner_gguf_base" placeholder="C:/models/base.gguf" title="Базовая GGUF для convert_lora_to_gguf" /></label>
<label>Рабочая директория <input type="text" id="sa_runner_workdir" placeholder="Assistent/training/runner" /></label>
<label>Custom command <input type="text" id="sa_runner_cmd" placeholder="{python} train.py --config {config}" /></label>
<label>convert_lora_to_gguf.py <input type="text" id="sa_runner_gguf_script" placeholder="путь к convert_lora_to_gguf.py" /></label>
<label>convert_lora_to_gguf.py <input type="text" id="sa_runner_gguf_script" placeholder="C:/llama.cpp/convert_lora_to_gguf.py" /></label>
<label>GGUF cmd template <input type="text" id="sa_runner_gguf_cmd" placeholder='"{python}" "{script}" "{base}" "{lora}" "{out}"' /></label>
<div class="sa-settings-row">
<button type="button" class="basic-button" id="sa_btn_refresh_models">Обновить модели</button>
<button type="button" class="basic-button" id="sa_btn_refresh_inventory">Обновить inventory</button>