Recipe catalog
Soup ships 181 ready-made configs at v0.78.0, and this page lists every one of them. The Recipes page explains how to use them; this one is the catalog, grouped by training task. v0.78.0 added eight recipes, see What changed in v0.78.0; v0.77.2 added two recipes, see What changed in v0.77.2; v0.77.0 added two recipes, removed one and marked two unavailable, see What changed in v0.77.0; v0.76.0 added nine, removed six and changed 38 more, see What changed in v0.76.0.
soup recipes list # all of them
soup recipes search qwen --task grpo # keyword, --task and --size filters
soup recipes show llama3.1-8b-sft # print the YAML
soup recipes use llama3.1-8b-sft -o soup.yaml # write it, then: soup train --config soup.yamlWhat a recipe is, and is not
A recipe is a complete soup.yaml for one base model and one task. At v0.78.0, 179 of the 181 load through the config schema with no unknown keys and no refusal, and the other two, kimi-k2-sft and kimi-k2-thinking-grpo, are refused when their config loads, and soup recipes use exits 3 on them (see What changed in v0.77.0 and Recipes): that was checked at v0.78.0 by loading every one through the schema, and by running soup recipes use and soup doctor --config on each of the eight new ones. That says the file is valid. It does not say the run works end to end, and it does not say it was trained to a quality target: the data paths are placeholders you replace, a few recipes stop at trainer setup or carry a key the trainer ignores (the list below names them, as checked at v0.77.0, plus the v0.78.0 additions), and this page loaded every recipe through the schema without training any of them. Treat every hyperparameter as a starting point, and read the base model's license before you train on it.
The Size column is the base model's total parameter count as the catalog states it (for a mixture-of-experts model, all resident parameters, not the active ones), not a memory figure. Since v0.76.0 a test compares the label of every recipe whose base has a recorded parameter count (all but four bases the Hub has no count for) with that count and allows 0.70x to 1.15x (#1161), so the labels are approximate rather than exact. Nothing checked them before that: through v0.75.2 deepseek-v3-7b-sft listed 7B for DeepSeek-V3-0324 (it now lists 685B, and the 7b in its name still does not describe the size), and glm-4.6-sft, glm-5-sft and minimax-m2-sft listed 9B. N/A means the catalog gives no size; the test pins those 13 recipes by name. Check the model card before you size a GPU. Quantization is the training.quantization the recipe sets (4bit is 4-bit NF4 loading, the QLoRA setup; none means no quantization; 8bit is 8-bit bitsandbytes loading; not set leaves the schema default, which is 4bit). bitnet_1.58 no longer appears: it is refused at config load and its only recipe was removed. Recipes that run on Apple Silicon say mlx in the notes.
What changed in v0.78.0
- Eight recipes added.
qwen3.5-4b-grpo(#2522),qwen3.5-27b-grpo(#2605),qwen3.5-122b-a10b-grpo(#2660),qwen3.5-397b-a17b-grpo(#2718),qwen3.5-122b-a10b-dpo(#2606),minimax-m3-dpo(#2719),qwen3.5-9b-pretrain(#2720) andqwen3.5-2b-pretrain(#2721), by @kok-o, @nazek52 and @abdumutalipabay0. Upstream says the Qwen3.5 and MiniMax recipes use inherited settings and are not tuned or measured (tracked as #275). None was trained for this page; all eight write withsoup recipes useand passsoup doctor --configat v0.78.0 (for these tasks that pass means "not reviewed": the command printsNone of the 0 setting(s) known to be unreadwhere it has no table). Nothing was removed or edited: the catalog was 173 and is 181. - What upstream checked, and what it did not. For seven of the eight (all but
minimax-m3-dpo), the pinned public configuration's text architecture and LoRA attachment were checked on the meta device, and a tiny random decoder's forward and backward pass on CPU. Forminimax-m3-dpoupstream's tests run one DPO step on a tiny LLaVA that has the wrapper's shape, not on the 428B weights, NF4 or a multi-GPU setup. For all eight, full-checkpoint training, NF4 execution, GPU memory use and training quality are not checked. The three large MoE recipes (qwen3.5-122b-a10b-dpo,qwen3.5-122b-a10b-grpo,qwen3.5-397b-a17b-grpo) say in the catalog that multi-GPU or multi-node memory fit is unverified, and upstream's docs add that no GPU count or memory-fit recommendation has been measured and that ten (seventeen) billion active parameters do not mean ten (seventeen) billion resident parameters. minimax-m3-dpoexists now. It is MiniMax-M3 preference training through DPO's vision path (modality: vision, attention-only LoRA on the language tower, text-onlyprompt,chosenandrejectedrows). The processor is custom code, so pass--trust-remote-code, and the MiniMax Community License needs a separate agreement for commercial use. (This site said through v0.77.2 that nominimax-m3-dporecipe was in the catalog; that was true then, and since v0.78.0 it is not.)- Search by size changed.
soup recipes search --sizematches the size a recipe declares, as a whole token (#2930, #2932): see Recipes.
What changed in v0.77.2
- Two recipes added.
qwen3.5-0.8b-dpo(#2184) andqwen3.5-27b-dpo(#2445), both by @kok-o, are text DPO recipes forQwen/Qwen3.5-0.8BandQwen/Qwen3.5-27B(modality: text, LoRA, 4-bit). Upstream says they use inherited Qwen 3.5 DPO settings, not hyperparameters tuned for either size, and that they are not tuned or measured (tracked as #275). The 0.8B setsbatch_size: auto; the 27B setsgradient_accumulation_steps: 8. None was trained for this page; both write withsoup recipes useand passsoup doctor --configat v0.77.2. - Nothing else changed. No recipe was removed or edited, and
kimi-k2-sftandkimi-k2-thinking-grpowere still the two that were refused at load. The catalog was 173, of which the DPO group was then 21.
What changed in v0.77.0
- Two recipes added.
qwen3.5-4b-dpoandqwen3.5-2b-dpoare text DPO recipes forQwen/Qwen3.5-4BandQwen/Qwen3.5-2B(modality: text, so the language decoder only). Upstream says their settings are inherited from the existing Qwen 3.5 DPO recipes, and it claims no training or performance result for either (#2129, #2160). None was trained for this page. - One recipe removed.
pixtral-dpo: itsllavarows had nochosen/rejectedcolumns, and v0.77.0 refusesllavaandsharegpt4vrows on preference tasks at config load (#1867, #2064).soup recipes show pixtral-dpoandsoup recipes use pixtral-dponow exit 1 withRecipe not found. - Two recipes marked unavailable.
kimi-k2-sftandkimi-k2-thinking-grpostay in the catalog, and at v0.77.0 a resident Kimi K2 load was refused at config load on Transformers 5.x (#2104, #2130).soup recipes showstill prints their YAML;soup recipes useprints the reason, writes nothing and exits 3. At v0.77.0 Kimi-K2-Base and -Instruct were admitted only undertraining.stream_layers: true, which upstream verified on CPU with a tiny K2-shaped checkpoint only (nothing has been trained on K2); Kimi-K2-Thinking has no supported route. (Since v0.78.0 a resident route for Kimi-K2-Base and -Instruct is admitted fortask: sftandtask: grpoatquantization: none, with the same thin evidence; the two recipes stay unavailable becausekimi-k2-sftsetsquantization: 4bitand Thinking has no route.) See Recipes. - 22 recipes gained
training.gradient_checkpointing: true. Since v0.77.0 the key reaches the TRL trainers, and with its defaultfalsethegrpo,ipo,bco,reward_modelandonline_dpotasks train with checkpointing off, where TRL's own default (on) used to apply. Everygrpo,ipo,reward_modelandonline_dporecipe in the catalog now sets it, so they keep the memory they had (#1923).
What changed in v0.76.0
- Nine recipes added.
kimi-k2.5-dpoandkimi-k2.5-grpo,deepseek-v4-pro-dpoanddeepseek-v4-pro-grpo,qwen3.5-35b-a3b-grpo,qwen3.6-35b-a3b-dpoandqwen3.6-35b-a3b-grpo,qwen3.6-27b-dpoandqwen3.6-27b-grpo. Upstream's changelog says the Kimi K2.5 pair is not trained and takes its hyperparameters from the K2.6 recipes. None of the nine was trained for this page either. - Six recipes removed.
soup recipes showandsoup recipes useof any of them now exit 1 withRecipe not found.cogito-v2-sftandgemma3-9b-sft: no compatible replacement base exists for their shipped definitions (#1122).mistral-large-3-sft: Mistral-Large-3 is published in Mistral's nativeparams.jsonlayout with noconfig.json, whichAutoConfigcannot read (#1398).falcon-e-bitnet-sft:quantization: bitnet_1.58is refused at config load because BitNet training is not implemented (#825).sesame-csm-ttsandspark-tts: their raw-audio configs are refused at config load (#265; since v0.77.0 Sesame CSM trains natively, #1769, and the recipe was not restored). See Removed in v0.76.0. - Seven base model IDs fixed.
devstral-sft,glm-4.6-sft,granite-4-sft,internvl-3-5-sft,kimi-k2-sft,mistral-medium-3-5-sftandvoxtral-sftnamed invalid Hugging Face IDs; the table below shows the IDs they use now (#1122). - Two recipes changed path.
minimax-m3-sftandmistral-medium-3-5-sftnow setmodality: vision, because their bases are vision-language wrappers.oute-ttsnow readsdata.format: chatml(pre-encoded codec tokens) instead of raw audio. - 23 existing recipes gained
lora.dropout: 0.0(22 MoE recipes that keepmoe_lora, andminimax-m3-sft, which kept the line when it droppedmoe_lora). peft refuses dropout on the fused 3-D expert parameters of a Qwen3-MoE under transformers 5.x, so with the schema default of0.05the LoRA attach fails (#798). If you turnmoe_loraon for a MoE base of your own, setlora.dropout: 0.0as these recipes do.
Before you run one
Loading as a config and running are different things. These are the recipes where the YAML is valid and the run still stops, needs something you have to supply first, or carries a setting that does not do what its name says. Each was checked against the v0.76.0 source, and the points below that name v0.77.0 were checked again at that tag; none was trained to confirm it.
soup recipes verify is new in v0.76.0 and checks one narrower thing for every recipe at once: whether the LoRA adapter the recipe asks for attaches to its base model. It builds the base's architecture on the meta device from its config.json, so no weights are downloaded, but it fetches each config.json from the Hub and needs the training stack installed. It does not look at your data, your memory or whether training converges, and its exit 0 can mean every config was unverified, so read the summary line. Details are on Recipes; soup recipes verify --config soup.yaml checks a config of your own before a long download.
- Sharding is a launch flag, not a recipe key. Where a description names DeepSpeed ZeRO-2 or ZeRO-3 or FSDP2 (for example
llama3.1-70b-sft,qwen2.5-72b-sft,gemma3-27b-sft,qwen3-32b-sft,qwen3.6-27b-dpo,qwen3.6-27b-grpo,deepseek-r1-32b-grpoandllama3-70b-fsdp2) or says multi-node (the DeepSeek V4 Pro, Kimi K2.5 and Kimi K2.6 recipes, among others), the YAML does not select it. Pass--deepspeed zero2,--deepspeed zero3or--fsdp full_shard, plus--gpus N, tosoup train, as the multi-GPU page shows; for more than one machine v0.76.0 adds--nodes,--node-rank,--master-addrand--master-port, see Multi-node launch.llama3-70b-fsdp2setsuse_fsdp2_compile: true, andsoup train --config soup.yamlwithout--fsdpexits 1. - Some recipes need something first.
llama3.1-8b-ppoloadstraining.reward_model: ./output_rm, a reward model you train beforehand.llama3.1-8b-longctxsetsuse_flash_attn: true, which exits 1 without CUDA and theflash-attnpackage.baichuan-sftsetstraining.hub: modelscope, so the base model is fetched through ModelScope into./.soup_hub_cache.minimax-m3-sftloads a processor that is custom code, so it needssoup train --trust-remote-code. - MoE recipes and
moe_lora. Through v0.75.2 only the SFT and pretrain trainers and the layer-streaming setup readmoe_lora. The eight DPO and GRPO MoE recipes in the v0.75.2 catalog (qwen3.5-35b-a3b-dpo,deepseek-v4-flash-dpo,glm-5.1-dpo,kimi-k2.6-dpo,qwen3-30b-a3b-reasoning-grpo,deepseek-v4-flash-grpo,glm-5.1-grpoandkimi-k2.6-grpo) set it anyway. Upstream says they got attention-only LoRA, not what their YAML said, and that on transformers 5.x (which v0.75.2 requires) peft refused the LoRA attach outright forqwen3_moe, so those runs stopped there (#798, #1074). If you ran one of them, re-run. Since v0.76.0 the DPO, GRPO and the other preference and RL trainers read it, so the 15 DPO and GRPO recipes that set it now pass it on, with limits upstream states: since #1421 expert adapters attach through peft's fused-parameter path (target_parameters) on theqwen3_moe,deepseek_v3,glm4_moe,olmoe,qwen2_moe,qwen3_5_moe_text,minimaxandmixtralmodel types,minimax-m3-sftis attention-only LoRA for a different reason (it no longer setsmoe_lora, and its YAML comment says it now targets no experts), the Kimi K2.x recipes are untested rather than known-good because no tiny stand-in for that configuration builds in the installed transformers, and upstream's checks ran on tiny random stand-ins on CPU, never on a real 30B-plus checkpoint.moe_aux_loss_coeff: 0.01in the DPO and GRPO recipes is the default value and at v0.77.0 is still read only by SFT, pretrain andtts: on a DPO or GRPO recipe it does nothing. (Since v0.77.0 a non-default value is refused at load onbackend: unslothand withstream_layers, which apply no MoE auxiliary loss either.) - A reward domain that does nothing.
deepseek-v3-reasoning,kimi-k2-thinking-grpoandqwq-32b-grposetverifiable_domain: mathbeside areward_fnthat is notverifiable, andverifiable_domainis read only by theverifiablereward (still so at v0.77.0), so it does nothing there. - Vision-language bases.
minimax-m3-sftandmistral-medium-3-5-sftsetmodality: vision, so SFT loads the vision-language wrapper and LoRA stays on the language tower. Theminimax-m3-sftYAML says in a comment that the vision collator supervises every token, prompt included, and thattrain_on_responses_onlyis not applied on this path.pixtral-dpowas not usable as shipped through v0.76.1, because itsdata.format: llavarows carry noprompt,chosenorrejected, which is what DPO reads (#1867). (Since closed in v0.77.0: the recipe was removed and vision row formats are refused on preference tasks at config load, #2064.) The GRPO trainer is still text-only at v0.77.0, sollama3.2-vision-grpodoes not use images. - Speech and BitNet. Three TTS recipes remain.
orpheus-tts-sftandllasa-ttssetdata.format: audio, the live-codec path (Orpheus needspip install snac, Llasa the[audio]extra);oute-ttstrains on pre-encodedchatmlrows, because a raw-audio Oute config is refused at config load (still so at v0.77.0). Sesame CSM trains natively since v0.77.0, but there is no Sesame CSM recipe; see Modality II. There is no BitNet recipe:falcon-e-bitnet-sftwas removed, and the export path (soup export --format tq1_0) is the BitNet path that works, see BitNet.whisper-large-v3-ftgoes through the generic audio SFT path; for Whisper speech-to-text usewhisper-large-v3-asr.deepseek-v3-pipelinevalidates its pipeline setting and then runs data-parallel. - The v0.78.0 recipes.
minimax-m3-dpoloads a custom processor, so it needssoup train --trust-remote-code, and setsmodality: vision(DPO builds the vision-language wrapper; the rows are text only). The GRPO recipes score with the built-inaccuracyreward, so every prompt needs a reference answer, either ananswercolumn or the final assistant message. Upstream's docs say GRPO promotes a configured batch size of 1 to at least 4 for four generations, so the 122B and 397B GRPO recipes are not one-completion workloads, and that the whole 122B (397B) base, the rollout state and the reference model matter, not only the active parameters. None of these was trained or fitted to a card by anyone here. These points come from upstream's docs and the YAML at the tag; the recipes were loaded and written, not run.
SFT: supervised fine-tuning
| Recipe | Base model | Size | Quantization | What it is |
|---|---|---|---|---|
baichuan-sft | baichuan-inc/Baichuan2-13B-Chat | 13B | 4bit | Baichuan 2 13B chat SFT |
biomistral-7b-sft | BioMistral/BioMistral-7B | 7B | 4bit | BioMistral 7B medical/biomedical domain SFT |
codellama-13b-sft | codellama/CodeLlama-13b-Instruct-hf | 13B | 4bit | Code Llama 13B code-specialist SFT |
codellama-70b-sft | codellama/CodeLlama-70b-Instruct-hf | 70B | 4bit | Code Llama 70B code-specialist SFT (multi-GPU) |
deepseek-ocr-sft | deepseek-ai/DeepSeek-OCR | N/A | 4bit | DeepSeek-OCR vision OCR SFT |
deepseek-r1-distill-llama-8b-sft | deepseek-ai/DeepSeek-R1-Distill-Llama-8B | 8B | 4bit | DeepSeek-R1-Distill Llama 8B reasoning SFT |
deepseek-r1-distill-qwen-1.5b-sft | deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | 1.5B | 4bit | DeepSeek-R1-Distill Qwen 1.5B reasoning SFT |
deepseek-r1-distill-qwen-7b-sft | deepseek-ai/DeepSeek-R1-Distill-Qwen-7B | 7B | 4bit | DeepSeek-R1-Distill Qwen 7B reasoning SFT |
deepseek-v3-7b-sft | deepseek-ai/DeepSeek-V3-0324 | 685B | 4bit | DeepSeek V3 SFT with MoE LoRA. The base is DeepSeek-V3-0324, so the 7b in the name does not describe the size |
deepseek-v3-pipeline | deepseek-ai/DeepSeek-V3 | 671B | 4bit | DeepSeek V3 SFT scaffold. Its parallelism: pipeline (4 stages) is checked (it needs CUDA and at least 4 GPUs) and then the run is data-parallel at v0.77.0, see Pipeline parallelism |
deepseek-v4-flash-sft | deepseek-ai/DeepSeek-V4-Flash | N/A | 4bit | DeepSeek V4 Flash MoE SFT (MIT, efficiency-tier) |
deepseek-v4-pro-sft | deepseek-ai/DeepSeek-V4-Pro | N/A | 4bit | DeepSeek V4 Pro flagship MoE SFT (MIT, 1.6T-class). Requires multi-node DeepSpeed. |
devstral-sft | mistralai/Devstral-Small-2507 | 24B | 4bit | Devstral Small code/agent SFT |
gemma2-2b-sft | google/gemma-2-2b-it | 2B | 4bit | Gemma 2 2B SFT (edge-friendly) |
gemma3-12b-sft | google/gemma-3-12b-it | 12B | 4bit | Gemma 3 12B instruction tuning |
gemma3-27b-sft | google/gemma-3-27b-it | 27B | 4bit | Gemma 3 27B SFT with DeepSpeed ZeRO-2 |
gemma3-4b-sft-mlx | mlx-community/gemma-3-4b-it-4bit | 4B | 4bit | Gemma 3 4B SFT on Apple Silicon via MLX (M1+ 16GB) (mlx) |
glm-4.6-sft | zai-org/GLM-4.6 | 357B | 4bit | GLM 4.6 MoE instruction tuning with LoRA |
glm-5-sft | zai-org/GLM-5 | 754B | 4bit | GLM 5 MoE SFT (next-gen GLM family) |
glm-5.1-sft | zai-org/GLM-5.1 | 754B | 4bit | GLM 5.1 MoE SFT (MIT, 754B). Multi-GPU / multi-node recommended. |
gpt-oss-120b-sft | openai/gpt-oss-120b | 120B | 4bit | GPT-OSS 120B SFT (multi-GPU recommended) |
gpt-oss-20b-sft | openai/gpt-oss-20b | 20B | 4bit | GPT-OSS 20B SFT (reasoning_effort=medium) |
granite-4-sft | ibm-granite/granite-4.0-h-tiny-base | 7B | 4bit | IBM Granite 4.0 tiny SFT |
internvl-2.5-8b-sft | OpenGVLab/InternVL2_5-8B | 8B | 4bit | InternVL 2.5 8B vision-language SFT |
internvl-3-5-sft | OpenGVLab/InternVL3_5-8B | 8B | 4bit | InternVL 3.5 vision SFT |
kimi-k2-sft | moonshotai/Kimi-K2-Base | N/A | 4bit | Unavailable since v0.77.0. The catalog now describes it as "Kimi K2 SFT: resident unsupported; guarded streaming only (see recipes use)." soup recipes use exits 3 and writes nothing: since v0.78.0 the message says the resident Kimi K2 route refuses training.quantization: 4bit, which this recipe sets (the route takes quantization: none; 4-bit goes through training.stream_layers: true). Evidence for either route is CPU parity on a tiny K2-shaped checkpoint; nothing trained on K2 |
kimi-k2.5-sft | moonshotai/Kimi-K2.5 | 1T | 4bit | Kimi K2.5 MoE SFT (Modified MIT, ~1T / 32B active). Requires multi-node DeepSpeed. |
kimi-k2.6-sft | moonshotai/Kimi-K2.6 | 1T | 4bit | Kimi K2.6 MoE SFT (Modified MIT, ~1T / 32B active). Requires multi-node DeepSpeed. |
lfm2-sft | LiquidAI/LFM2-1.2B | 1.2B | 4bit | Liquid LFM2 1.2B SFT (edge-optimised) |
llama2-13b-finance-sft | meta-llama/Llama-2-13b-hf | 13B | 4bit | Llama 2 13B finance-domain SFT (FinGPT-style starter recipe) |
llama3-70b-fsdp2 | meta-llama/Llama-3.1-70B-Instruct | 70B | 4bit | Llama 3.1 70B SFT with FSDP2 full shard + torch.compile. Requires 8 x A100/H100 80GB. |
llama3.1-70b-sft | meta-llama/Llama-3.1-70B-Instruct | 70B | 4bit | Llama 3.1 70B SFT with DeepSpeed ZeRO-3 |
llama3.1-8b-longctx | meta-llama/Llama-3.1-8B-Instruct | 8B | 4bit | Llama 3.1 8B long-context (32k) with YaRN RoPE scaling |
llama3.1-8b-sft | meta-llama/Llama-3.1-8B-Instruct | 8B | 4bit | Llama 3.1 8B instruction tuning with LoRA |
llama3.1-8b-sft-mlx | mlx-community/Llama-3.1-8B-Instruct-4bit | 8B | 4bit | Llama 3.1 8B SFT on Apple Silicon via MLX (M2+ 16GB) (mlx) |
llama3.2-11b-vision | meta-llama/Llama-3.2-11B-Vision-Instruct | 11B | 4bit | Llama 3.2 11B Vision multimodal fine-tuning |
llama3.2-1b-sft | meta-llama/Llama-3.2-1B-Instruct | 1B | 8bit | Llama 3.2 1B instruction tuning (mobile-friendly) |
llama3.2-3b-sft | meta-llama/Llama-3.2-3B-Instruct | 3B | 4bit | Llama 3.2 3B instruction tuning (edge-friendly) |
llama3.2-vision-90b-sft | meta-llama/Llama-3.2-90B-Vision-Instruct | 90B | 4bit | Llama 3.2 90B Vision multimodal SFT (8 x A100/H100 80GB) |
llama4-scout-17b-sft | meta-llama/Llama-4-Scout-17B-16E-Instruct | 109B | 4bit | Llama 4 Scout 17B active SFT with LoRA (4bit) |
llama4-scout-tools | meta-llama/Llama-4-Scout-17B-16E-Instruct | 109B | 4bit | Llama 4 Scout 17B active tool-calling / function-calling SFT |
llava-next-sft | llava-hf/llava-v1.6-mistral-7b-hf | 7B | 4bit | LLaVA-Next 7B vision SFT |
magicoder-7b-sft | ise-uiuc/Magicoder-S-DS-6.7B | 6.7B | 4bit | Magicoder S-DS 6.7B code-specialist SFT |
magistral-small-sft | mistralai/Magistral-Small | 24B | 4bit | Magistral Small reasoning SFT |
mathstral-7b-sft | mistralai/Mathstral-7B-v0.1 | 7B | 4bit | Mathstral 7B math/STEM-specialist SFT |
medgemma-sft | google/medgemma-4b-it | 4B | 4bit | MedGemma 4B medical SFT |
meditron-7b-sft | epfl-llm/meditron-7b | 7B | 4bit | Meditron 7B medical / clinical domain SFT |
minicpm-v-2.6-sft | openbmb/MiniCPM-V-2_6 | 8B | 4bit | MiniCPM-V 2.6 vision-language SFT (edge-friendly multimodal) |
minimax-m2-sft | MiniMaxAI/MiniMax-M2 | 230B | 4bit | MiniMax M2 MoE SFT instruction tuning |
minimax-m3-sft | MiniMaxAI/MiniMax-M3 | 428B | 4bit | MiniMax M3 MoE SFT (428B / 23B active). MiniMax Community License - commercial use requires a separate agreement. Multi-GPU recommended. Sets modality: vision, so SFT loads the vision-language wrapper, and its processor is custom code: run soup train --trust-remote-code. It no longer sets moe_lora, so no expert is adapted; target_modules: auto attaches to the language tower alone. The recipe's own comment says the vision collator supervises every token, prompt included, so train_on_responses_only is not applied on this path |
ministral-sft | mistralai/Ministral-8B-Instruct-2410 | 8B | 4bit | Ministral 8B SFT |
mistral-7b-sft | mistralai/Mistral-7B-Instruct-v0.3 | 7B | 4bit | Mistral 7B instruction tuning |
mistral-medium-3-5-sft | mistralai/Mistral-Medium-3.5-128B | 128B | 4bit | Mistral Medium 3.5 SFT. Sets modality: vision, so SFT loads the vision-language wrapper (Mistral3ForConditionalGeneration) and target_modules: auto reaches the language model's attention projections, not the vision tower (#1395) |
mistral-small-3-sft | mistralai/Mistral-Small-24B-Instruct-2501 | 24B | 4bit | Mistral Small 3 24B SFT |
nemotron-4-340b-sft | nvidia/Nemotron-4-340B-Instruct | 340B | 4bit | Nemotron-4 340B SFT (massive multi-node deployment) |
paddle-ocr-sft | PaddlePaddle/PaddleOCR-VL | N/A | 4bit | Paddle-OCR-VL OCR SFT |
phi3.5-mini-sft | microsoft/Phi-3.5-mini-instruct | 3.8B | 4bit | Phi-3.5-mini 3.8B SFT (small / edge) |
phi4-14b-sft | microsoft/phi-4 | 14B | 4bit | Phi-4 14B instruction tuning |
pixtral-12b-sft | mistralai/Pixtral-12B-2409 | 12B | 4bit | Pixtral 12B vision-language SFT with LoRA |
qvq-72b-sft | Qwen/QVQ-72B-Preview | 72B | 4bit | QVQ 72B vision-reasoning SFT |
qwen-image-sft | Qwen/Qwen-Image | N/A | 4bit | Qwen-Image image-output multimodal SFT |
qwen2-audio-7b-sft | Qwen/Qwen2-Audio-7B-Instruct | 7B | 4bit | Qwen2-Audio 7B audio-language SFT |
qwen2-vl-72b-sft | Qwen/Qwen2-VL-72B-Instruct | 72B | 4bit | Qwen2-VL 72B vision-language SFT (multi-GPU recommended) |
qwen2-vl-7b-sft | Qwen/Qwen2-VL-7B-Instruct | 7B | 4bit | Qwen2-VL 7B vision-language SFT |
qwen2.5-0.5b-sft | Qwen/Qwen2.5-0.5B-Instruct | 0.5B | 8bit | Qwen 2.5 0.5B SFT (mobile / edge) |
qwen2.5-1.5b-sft | Qwen/Qwen2.5-1.5B-Instruct | 1.5B | 8bit | Qwen 2.5 1.5B SFT (edge-friendly) |
qwen2.5-3b-sft | Qwen/Qwen2.5-3B-Instruct | 3B | 4bit | Qwen 2.5 3B SFT (small / edge) |
qwen2.5-72b-sft | Qwen/Qwen2.5-72B-Instruct | 72B | 4bit | Qwen 2.5 72B SFT with DeepSpeed ZeRO-3 |
qwen2.5-7b-sft | Qwen/Qwen2.5-7B-Instruct | 7B | 4bit | Qwen 2.5 7B instruction tuning with LoRA |
qwen2.5-coder-1.5b-sft | Qwen/Qwen2.5-Coder-1.5B-Instruct | 1.5B | 4bit | Qwen 2.5 Coder 1.5B instruction tuning with LoRA |
qwen2.5-coder-14b-sft | Qwen/Qwen2.5-Coder-14B-Instruct | 14B | 4bit | Qwen 2.5 Coder 14B instruction tuning with LoRA |
qwen2.5-coder-32b-sft | Qwen/Qwen2.5-Coder-32B-Instruct | 32B | 4bit | Qwen 2.5 Coder 32B instruction tuning with LoRA |
qwen2.5-coder-7b-sft | Qwen/Qwen2.5-Coder-7B-Instruct | 7B | 4bit | Qwen 2.5 Coder 7B instruction tuning with LoRA |
qwen2.5-math-1.5b-sft | Qwen/Qwen2.5-Math-1.5B-Instruct | 1.5B | 4bit | Qwen 2.5 Math 1.5B instruction tuning with LoRA |
qwen2.5-math-7b-sft | Qwen/Qwen2.5-Math-7B-Instruct | 7B | 4bit | Qwen 2.5 Math 7B instruction tuning with LoRA |
qwen3-14b-sft | Qwen/Qwen3-14B | 14B | 4bit | Qwen 3 14B instruction tuning with LoRA |
qwen3-30b-a3b-sft | Qwen/Qwen3-30B-A3B | 30B | 4bit | Qwen 3 30B-A3B MoE instruction tuning |
qwen3-32b-sft | Qwen/Qwen3-32B | 32B | 4bit | Qwen 3 32B SFT with DeepSpeed ZeRO-2 |
qwen3-32b-zeropp | Qwen/Qwen3-32B | 32B | 4bit | Qwen3 32B SFT with DeepSpeed ZeRO++ (quantized gradients + hierarchical partitioning). Launch with --deepspeed zero++. |
qwen3-8b-sft | Qwen/Qwen3-8B | 8B | 4bit | Qwen 3 8B instruction tuning |
qwen3-8b-sft-mlx | mlx-community/Qwen3-8B-4bit | 8B | 4bit | Qwen 3 8B SFT on Apple Silicon via MLX (M2+ 16GB) (mlx) |
qwen3-8b-tools | Qwen/Qwen3-8B | 8B | 4bit | Qwen 3 8B tool-calling / function-calling SFT |
qwen3-coder-30b-sft | Qwen/Qwen3-Coder-30B-A3B-Instruct | 30B | 4bit | Qwen3-Coder 30B (A3B MoE) code-specialist SFT |
qwen3.5-0.8b-sft | Qwen/Qwen3.5-0.8B | 0.8B | 4bit | Qwen 3.5 0.8B SFT (Apache-2.0, tiny / mobile) |
qwen3.5-122b-a10b-sft | Qwen/Qwen3.5-122B-A10B | 122B | 4bit | Qwen 3.5 122B-A10B MoE SFT (Apache-2.0, 10B active). Multi-GPU recommended (8 x A100/H100 80GB). |
qwen3.5-27b-sft | Qwen/Qwen3.5-27B | 27B | 4bit | Qwen 3.5 27B SFT (Apache-2.0) with DeepSpeed ZeRO-2 |
qwen3.5-2b-sft | Qwen/Qwen3.5-2B | 2B | 4bit | Qwen 3.5 2B SFT (Apache-2.0, small / edge) |
qwen3.5-35b-a3b-sft | Qwen/Qwen3.5-35B-A3B | 35B | 4bit | Qwen 3.5 35B-A3B MoE SFT (Apache-2.0, 3B active) |
qwen3.5-397b-a17b-sft | Qwen/Qwen3.5-397B-A17B | 397B | 4bit | Qwen 3.5 397B-A17B flagship MoE SFT (Apache-2.0, 17B active). Requires multi-node DeepSpeed / FSDP. |
qwen3.5-4b-sft | Qwen/Qwen3.5-4B | 4B | 4bit | Qwen 3.5 4B SFT (Apache-2.0, 262K context) |
qwen3.5-9b-sft | Qwen/Qwen3.5-9B | 9B | 4bit | Qwen 3.5 9B SFT (Apache-2.0, 262K context) |
qwen3.6-27b-sft | Qwen/Qwen3.6-27B | 27B | 4bit | Qwen 3.6 27B SFT (Apache-2.0) with DeepSpeed ZeRO-2 |
qwen3.6-35b-a3b-sft | Qwen/Qwen3.6-35B-A3B | 35B | 4bit | Qwen 3.6 35B-A3B MoE SFT (Apache-2.0, 3B active) |
qwen3.8-27b-sft | Qwen/Qwen3.8-27B | 27B | 4bit | Qwen 3.8 27B text-only SFT (Apache-2.0) with DeepSpeed ZeRO-2 |
ra-dit-llama3-8b | meta-llama/Llama-3.1-8B-Instruct | 8B | 4bit | RA-DIT stage 2 (Meta 2023) — RAFT-style SFT on the generator. Pairs with ra-dit-retriever (stage 1). Uses RAFT data format. |
raft-llama3-8b | meta-llama/Llama-3.1-8B-Instruct | 8B | 4bit | RAFT (Retrieval-Augmented Fine-Tuning, Stanford 2024) — train an 8B Llama 3.1 to answer queries given a golden doc + distractor docs. Rows: {query, golden_doc, distractor_docs, answer}. |
seamlessm4t-v2-sft | facebook/seamless-m4t-v2-large | 2.3B | 4bit | SeamlessM4T v2 multilingual speech-to-text SFT |
smollm2-1.7b-sft | HuggingFaceTB/SmolLM2-1.7B-Instruct | 1.7B | 8bit | SmolLM2 1.7B SFT (small / edge) |
smollm2-135m-sft | HuggingFaceTB/SmolLM2-135M-Instruct | 135M | none | SmolLM2 135M SFT (ultra-tiny / mobile) |
smollm2-360m-sft | HuggingFaceTB/SmolLM2-360M-Instruct | 360M | none | SmolLM2 360M SFT (tiny / mobile) |
smollm3-3b-sft | HuggingFaceTB/SmolLM3-3B | 3B | 8bit | SmolLM3 3B SFT (small / edge) |
smolvlm-256m-sft | HuggingFaceTB/SmolVLM-256M-Instruct | 256M | none | SmolVLM 256M vision SFT (llava format) — a tiny VLM. SmolVLM uses an Idefics3 processor; Soup mirrors its nested tokenizer surface and performs processor-aware vision collation so image-token expansion and pixel_values reach the model. target_modules are pinned to q_proj/v_proj (auto cannot infer them for Idefics3). |
voxtral-sft | mistralai/Voxtral-Mini-3B-2507 | 4.7B | 4bit | Voxtral Mini 3B audio SFT |
whisper-large-v3-ft | openai/whisper-large-v3 | 1.5B | 8bit | Whisper Large v3 through the generic audio SFT path (modality: audio), which loads the model with AutoModel, the bare Whisper encoder-decoder without a language-model head. For Whisper speech-to-text use the ASR task, see ASR fine-tuning and whisper-large-v3-asr below |
DPO: preference optimization
| Recipe | Base model | Size | Quantization | What it is |
|---|---|---|---|---|
deepseek-r1-distill-llama-8b-dpo | deepseek-ai/DeepSeek-R1-Distill-Llama-8B | 8B | 4bit | DeepSeek-R1-Distill Llama 8B DPO alignment |
deepseek-r1-distill-qwen-1.5b-dpo | deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | 1.5B | 4bit | DeepSeek-R1-Distill Qwen 1.5B DPO alignment |
deepseek-r1-distill-qwen-7b-dpo | deepseek-ai/DeepSeek-R1-Distill-Qwen-7B | 7B | 4bit | DeepSeek-R1-Distill Qwen 7B DPO alignment |
deepseek-v4-flash-dpo | deepseek-ai/DeepSeek-V4-Flash | N/A | 4bit | DeepSeek V4 Flash MoE DPO alignment (MIT, efficiency-tier) |
deepseek-v4-pro-dpo | deepseek-ai/DeepSeek-V4-Pro | N/A | 4bit | DeepSeek V4 Pro MoE DPO alignment (MIT, 1.6T-class). Sets moe_lora: true. The catalog says it requires multi-node DeepSpeed, which is a launch flag, see Before you run one |
gemma3-27b-dpo | google/gemma-3-27b-it | 27B | 4bit | Gemma 3 27B DPO alignment |
glm-5.1-dpo | zai-org/GLM-5.1 | 754B | 4bit | GLM 5.1 MoE DPO alignment (MIT, 754B). Multi-GPU / multi-node recommended. |
kimi-k2.5-dpo | moonshotai/Kimi-K2.5 | 1T | 4bit | Kimi K2.5 MoE DPO alignment (Modified MIT, ~1T / 32B active). Sets moe_lora: true. The catalog says it requires multi-node DeepSpeed; its hyperparameters are inherited from the Kimi K2.6 recipe, not tuned for K2.5 |
kimi-k2.6-dpo | moonshotai/Kimi-K2.6 | 1T | 4bit | Kimi K2.6 MoE DPO alignment (Modified MIT, ~1T / 32B active). Requires multi-node DeepSpeed. |
llama3.1-8b-dpo | meta-llama/Llama-3.1-8B-Instruct | 8B | 4bit | Llama 3.1 8B DPO alignment |
llama4-scout-17b-dpo | meta-llama/Llama-4-Scout-17B-16E-Instruct | 109B | 4bit | Llama 4 Scout 17B active DPO alignment |
minimax-m3-dpo | MiniMaxAI/MiniMax-M3 | 428B | 4bit | New in v0.78.0 (#2719, by @abdumutalipabay0). MiniMax M3 MoE DPO alignment (428B / 23B active), a vision-language checkpoint trained through DPO's vision path on text rows (modality: vision, LoRA r 32 / alpha 64, lora.dropout: 0.0, batch 1, 16 accumulation steps); the processor is custom code, so pass --trust-remote-code. MiniMax Community License: commercial use needs a separate agreement. Multi-GPU recommended. Inherited settings, not tuned or measured (#275); not trained here |
mistral-7b-dpo | mistralai/Mistral-7B-Instruct-v0.3 | 7B | 4bit | Mistral 7B DPO alignment |
qwen2.5-7b-dpo | Qwen/Qwen2.5-7B-Instruct | 7B | 4bit | Qwen 2.5 7B DPO alignment |
qwen3.5-0.8b-dpo | Qwen/Qwen3.5-0.8B | 0.8B | 4bit | Qwen 3.5 0.8B text DPO alignment (Apache-2.0, tiny / mobile). New in v0.77.2 (#2184, by @kok-o); sets modality: text and batch_size: auto. Settings inherited from the other Qwen 3.5 DPO recipes, not tuned or measured for 0.8B (#275) |
qwen3.5-122b-a10b-dpo | Qwen/Qwen3.5-122B-A10B | 122B | 4bit | New in v0.78.0 (#2606, by @kok-o). Qwen 3.5 122B-A10B MoE text DPO alignment (Apache-2.0, 10B active). Sets modality: text, moe_lora: true, lora.dropout: 0.0, LoRA r 32 / alpha 64, batch 1, 16 accumulation steps and dpo_beta: 0.1. Inherited starting settings; the catalog says multi-GPU memory fit is unverified. Not trained here |
qwen3.5-27b-dpo | Qwen/Qwen3.5-27B | 27B | 4bit | Qwen 3.5 27B text DPO alignment (Apache-2.0). New in v0.77.2 (#2445, by @kok-o); sets modality: text and gradient_accumulation_steps: 8. Settings inherited from the other Qwen 3.5 DPO recipes, not tuned or measured (#275) |
qwen3.5-2b-dpo | Qwen/Qwen3.5-2B | 2B | 4bit | Qwen 3.5 2B text DPO alignment (Apache-2.0, small / edge). New in v0.77.0; sets modality: text, so it trains the language decoder on text preference pairs. Settings inherited from the other Qwen 3.5 DPO recipes; no training result is claimed |
qwen3.5-35b-a3b-dpo | Qwen/Qwen3.5-35B-A3B | 35B | 4bit | Qwen 3.5 35B-A3B MoE DPO alignment (Apache-2.0, 3B active) |
qwen3.5-4b-dpo | Qwen/Qwen3.5-4B | 4B | 4bit | Qwen 3.5 4B text DPO alignment (Apache-2.0). New in v0.77.0; sets modality: text. Settings inherited from the 9B DPO recipe; no training result is claimed |
qwen3.5-9b-dpo | Qwen/Qwen3.5-9B | 9B | 4bit | Qwen 3.5 9B DPO alignment (Apache-2.0) |
qwen3.6-27b-dpo | Qwen/Qwen3.6-27B | 27B | 4bit | Qwen 3.6 27B DPO alignment (Apache-2.0). The catalog says DeepSpeed ZeRO-2, which is a launch flag, not a key in the YAML. Sets modality: text |
qwen3.6-35b-a3b-dpo | Qwen/Qwen3.6-35B-A3B | 35B | 4bit | Qwen 3.6 35B-A3B MoE DPO alignment (Apache-2.0, 3B active). Sets moe_lora: true and modality: text |
GRPO: reinforcement learning with a verifier
| Recipe | Base model | Size | Quantization | What it is |
|---|---|---|---|---|
deepseek-r1-32b-grpo | deepseek-ai/DeepSeek-R1-Distill-Qwen-32B | 32B | 4bit | DeepSeek R1 32B GRPO with DeepSpeed |
deepseek-r1-8b-grpo | deepseek-ai/DeepSeek-R1-Distill-Llama-8B | 8B | 4bit | DeepSeek R1 8B GRPO reasoning |
deepseek-v3-reasoning | deepseek-ai/DeepSeek-V3 | N/A | 4bit | DeepSeek V3 (671B-parameter MoE) GRPO reasoning recipe. Requires multi-node DeepSpeed. Scores with the accuracy and format rewards; its verifiable_domain: math is only read by reward_fn: verifiable and has no effect here |
deepseek-v4-flash-grpo | deepseek-ai/DeepSeek-V4-Flash | N/A | 4bit | DeepSeek V4 Flash MoE GRPO reasoning training (MIT, efficiency-tier) |
deepseek-v4-pro-grpo | deepseek-ai/DeepSeek-V4-Pro | N/A | 4bit | DeepSeek V4 Pro MoE GRPO reasoning training (MIT, 1.6T-class). Sets moe_lora: true and reward_fn: accuracy. The catalog says it requires multi-node DeepSpeed, which is a launch flag |
glm-5.1-grpo | zai-org/GLM-5.1 | 754B | 4bit | GLM 5.1 MoE GRPO reasoning training (MIT, 754B). Multi-GPU / multi-node recommended. |
grpo-env-calculator | HuggingFaceTB/SmolLM2-135M-Instruct | 135M | not set | SmolLM2-135M GRPO on the bundled calculator env (openenv rollout) |
grpo-env-guess-number | HuggingFaceTB/SmolLM2-135M-Instruct | 135M | not set | SmolLM2-135M GRPO on the bundled number-deduction env (openenv rollout) |
grpo-env-retrieval-qa | HuggingFaceTB/SmolLM2-135M-Instruct | 135M | not set | SmolLM2-135M GRPO on the bundled retrieval-QA env (openenv rollout) |
kimi-k2-thinking-grpo | moonshotai/Kimi-K2-Thinking | N/A | 4bit | Unavailable since v0.77.0. The catalog now describes it as "Kimi K2 Thinking GRPO: unsupported resident route; no supported route yet." soup recipes use exits 3 and writes nothing; Kimi-K2-Thinking has no supported route |
kimi-k2.5-grpo | moonshotai/Kimi-K2.5 | 1T | 4bit | Kimi K2.5 MoE GRPO reasoning (Modified MIT, ~1T / 32B active). Sets moe_lora: true and reward_fn: accuracy. The catalog says it requires multi-node DeepSpeed; its hyperparameters are inherited from the Kimi K2.6 recipe, not tuned for K2.5 |
kimi-k2.6-grpo | moonshotai/Kimi-K2.6 | 1T | 4bit | Kimi K2.6 MoE GRPO reasoning (Modified MIT, ~1T / 32B active). Requires multi-node DeepSpeed. |
llama3.1-8b-grpo | meta-llama/Llama-3.1-8B-Instruct | 8B | 4bit | Llama 3.1 8B GRPO reasoning training |
llama3.2-vision-grpo | meta-llama/Llama-3.2-11B-Vision-Instruct | 11B | 4bit | Llama 3.2 11B Vision GRPO. The GRPO trainer is text-only at v0.77.0, so the images are not used, see GRPO Plus |
llama4-scout-17b-grpo | meta-llama/Llama-4-Scout-17B-16E-Instruct | 109B | 4bit | Llama 4 Scout 17B active GRPO reasoning training |
phi4-reasoning-grpo | microsoft/phi-4 | 14B | 4bit | Phi-4 14B GRPO reasoning training |
qwen2.5-7b-grpo | Qwen/Qwen2.5-7B-Instruct | 7B | 4bit | Qwen 2.5 7B GRPO reasoning training |
qwen3-30b-a3b-reasoning-grpo | Qwen/Qwen3-30B-A3B | 30B | 4bit | Qwen3 30B-A3B GRPO reasoning training (MoE thinking model) |
qwen3-8b-grpo | Qwen/Qwen3-8B | 8B | 4bit | Qwen 3 8B GRPO reasoning training |
qwen3.5-0.8b-grpo | Qwen/Qwen3.5-0.8B | 0.8B | 4bit | Qwen 3.5 0.8B GRPO reasoning training (Apache-2.0, tiny / mobile) |
qwen3.5-122b-a10b-grpo | Qwen/Qwen3.5-122B-A10B | 122B | 4bit | New in v0.78.0 (#2660, by @nazek52). Qwen 3.5 122B-A10B text MoE GRPO (Apache-2.0, 10B active). Sets moe_lora: true, reward_fn: accuracy, num_generations: 4, batch 1 and 16 accumulation steps; each prompt needs a reference answer (an answer column or the final assistant message). The catalog calls it an untuned multi-GPU starting configuration with unverified memory fit. Not trained here |
qwen3.5-2b-grpo | Qwen/Qwen3.5-2B | 2B | 4bit | Qwen 3.5 2B GRPO reasoning training (Apache-2.0, small / edge) |
qwen3.5-27b-grpo | Qwen/Qwen3.5-27B | 27B | 4bit | New in v0.78.0 (#2605, by @kok-o). Qwen 3.5 27B text GRPO reasoning training (Apache-2.0). Same settings as the 4B recipe (batch_size: auto, 8 accumulation steps, four generations, accuracy reward, gradient checkpointing). Inherited, not tuned or measured; not trained here |
qwen3.5-397b-a17b-grpo | Qwen/Qwen3.5-397B-A17B | 397B | 4bit | New in v0.78.0 (#2718, by @abdumutalipabay0). Qwen 3.5 397B-A17B text MoE GRPO (Apache-2.0, 17B active). Sets moe_lora: true, reward_fn: accuracy, batch 1 and 32 accumulation steps. The catalog calls it an untuned multi-node starting configuration with unverified memory fit; its SFT sibling already needs multi-node DeepSpeed or FSDP. Not trained here |
qwen3.5-4b-grpo | Qwen/Qwen3.5-4B | 4B | 4bit | New in v0.78.0 (#2522, by @kok-o). Qwen 3.5 4B text GRPO reasoning training (Apache-2.0). Inherits the 0.8B, 2B and 9B GRPO starting settings: batch_size: auto, 8 accumulation steps, four generations, reward_fn: accuracy, gradient checkpointing, 4096 tokens. Each prompt needs a reference answer. Not trained here |
qwen3.5-35b-a3b-grpo | Qwen/Qwen3.5-35B-A3B | 35B | 4bit | Qwen 3.5 35B-A3B MoE GRPO reasoning training (Apache-2.0, 3B active). Completes the SFT, DPO and GRPO set for this base. Sets moe_lora: true and modality: text |
qwen3.5-9b-grpo | Qwen/Qwen3.5-9B | 9B | 4bit | Qwen 3.5 9B GRPO reasoning training (Apache-2.0, 262K context) |
qwen3.6-27b-grpo | Qwen/Qwen3.6-27B | 27B | 4bit | Qwen 3.6 27B GRPO reasoning training (Apache-2.0). The catalog says DeepSpeed ZeRO-2, which is a launch flag, not a key in the YAML. Sets reward_fn: accuracy and modality: text |
qwen3.6-35b-a3b-grpo | Qwen/Qwen3.6-35B-A3B | 35B | 4bit | Qwen 3.6 35B-A3B MoE GRPO reasoning training (Apache-2.0, 3B active). Sets moe_lora: true and modality: text |
qwq-32b-grpo | Qwen/QwQ-32B | 32B | 4bit | QwQ 32B GRPO reasoning training |
r1-distill-llama-70b-grpo | deepseek-ai/DeepSeek-R1-Distill-Llama-70B | 70B | 4bit | DeepSeek-R1-Distill Llama 70B GRPO reasoning (multi-GPU) |
r1-distill-qwen-1.5b-grpo | deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B | 1.5B | 4bit | DeepSeek-R1-Distill Qwen 1.5B GRPO reasoning training |
r1-distill-qwen-14b-grpo | deepseek-ai/DeepSeek-R1-Distill-Qwen-14B | 14B | 4bit | DeepSeek-R1-Distill Qwen 14B GRPO reasoning training |
r1-distill-qwen-7b-grpo | deepseek-ai/DeepSeek-R1-Distill-Qwen-7B | 7B | 4bit | DeepSeek-R1-Distill Qwen 7B GRPO reasoning training |
KTO
| Recipe | Base model | Size | Quantization | What it is |
|---|---|---|---|---|
llama3.1-8b-kto | meta-llama/Llama-3.1-8B-Instruct | 8B | 4bit | Llama 3.1 8B KTO unpaired preference alignment |
ORPO
| Recipe | Base model | Size | Quantization | What it is |
|---|---|---|---|---|
llama3.1-8b-orpo | meta-llama/Llama-3.1-8B-Instruct | 8B | 4bit | Llama 3.1 8B ORPO reference-free alignment |
SimPO
| Recipe | Base model | Size | Quantization | What it is |
|---|---|---|---|---|
llama3.1-8b-simpo | meta-llama/Llama-3.1-8B-Instruct | 8B | 4bit | Llama 3.1 8B SimPO length-normalized alignment |
IPO
| Recipe | Base model | Size | Quantization | What it is |
|---|---|---|---|---|
llama3.1-8b-ipo | meta-llama/Llama-3.1-8B-Instruct | 8B | 4bit | Llama 3.1 8B IPO regularized preference alignment |
PPO
| Recipe | Base model | Size | Quantization | What it is |
|---|---|---|---|---|
llama3.1-8b-ppo | meta-llama/Llama-3.1-8B-Instruct | 8B | 4bit | Llama 3.1 8B PPO (RLHF stage 3) |
Online DPO
| Recipe | Base model | Size | Quantization | What it is |
|---|---|---|---|---|
online-dpo-smollm2-135m | HuggingFaceTB/SmolLM2-135M-Instruct | 135M | none | SmolLM2 135M Online DPO — on-policy generation judged by a pairwise LLM judge (the recipe sets training.online_dpo_judge: ollama://llama3.1, which needs a local Ollama server; point it at your own judge) |
Reward model
| Recipe | Base model | Size | Quantization | What it is |
|---|---|---|---|---|
qwen2.5-7b-reward | Qwen/Qwen2.5-7B-Instruct | 7B | 4bit | Qwen 2.5 7B reward model (RLHF stage 2) |
Embedding
| Recipe | Base model | Size | Quantization | What it is |
|---|---|---|---|---|
embedding-gemma-sft | google/embeddinggemma-300m | 300M | 4bit | EmbeddingGemma 300M sentence-embedding SFT |
llama3.1-8b-embed | meta-llama/Llama-3.1-8B | 8B | 4bit | Llama 3.1 8B sentence embedding with cosine loss |
ra-dit-retriever | sentence-transformers/all-mpnet-base-v2 | N/A | not set | RA-DIT stage 1 (Meta 2023) — contrastive retriever training. Pairs with ra-dit-llama3-8b (stage 2) for the full pipeline. |
Continued pre-training
| Recipe | Base model | Size | Quantization | What it is |
|---|---|---|---|---|
qwen2.5-7b-pretrain | Qwen/Qwen2.5-7B | 7B | 4bit | Qwen 2.5 7B continued pre-training |
qwen3.5-2b-pretrain | Qwen/Qwen3.5-2B-Base | 2B | 4bit | New in v0.78.0 (#2721, by @abdumutalipabay0). Qwen 3.5 2B Base continued pre-training (Apache-2.0) on plaintext rows, with the qwen3.5-4b-pretrain settings, inherited and not tuned for 2B. Not trained here |
qwen3.5-4b-pretrain | Qwen/Qwen3.5-4B-Base | 4B | 4bit | Qwen 3.5 4B Base continued pre-training (Apache-2.0) |
qwen3.5-9b-pretrain | Qwen/Qwen3.5-9B-Base | 9B | 4bit | New in v0.78.0 (#2720, by @abdumutalipabay0). Qwen 3.5 9B Base continued pre-training (Apache-2.0) on plaintext rows. Takes the qwen3.5-4b-pretrain settings (one epoch, lr: 1e-4, batch_size: auto, 8 accumulation steps, LoRA r 32 / alpha 64, 4096 tokens), inherited and not tuned for 9B. Not trained here |
ASR: speech recognition
| Recipe | Base model | Size | Quantization | What it is |
|---|---|---|---|---|
whisper-base-asr | openai/whisper-base | 74M | none | Whisper base (74M) ASR fine-tune with LoRA on q/v projections — fits a 4 GB GPU |
whisper-large-v3-asr | openai/whisper-large-v3 | 1.5B | none | Whisper large-v3 (1.5B) ASR fine-tune with LoRA — needs a larger GPU (>= 16 GB); the tiny/base recipes fit a 4 GB card |
whisper-tiny-asr | openai/whisper-tiny | 39M | none | Whisper tiny (39M) ASR fine-tune with LoRA on q/v projections — fits a 4 GB GPU. Rows: {"audio": path, "text": transcript} |
TTS: text to speech
| Recipe | Base model | Size | Quantization | What it is |
|---|---|---|---|---|
llasa-tts | HKUSTAudio/Llasa-1B | 1B | not set | Llasa-TTS. Sets data.format: audio, the live-codec path: raw audio is encoded with XCodec2 at training time and needs pip install "soup-cli[audio]", see TTS |
orpheus-tts-sft | canopylabs/orpheus-3b-0.1-ft | 3B | not set | Orpheus emotional TTS. Its data.format: audio is the live-codec path and needs pip install snac; the pre-encoded path needs no extra package, see TTS |
oute-tts | OuteAI/OuteTTS-0.3-500M | 0.5B | not set | OuteTTS v0.3 on pre-encoded data. Sets data.format: chatml, so your rows must already be codec tokens as text; a raw-audio (data.format: audio) Oute config is refused at config load, see TTS |
See also
- Recipes: how to list, search, show and use them, and
soup recipes verifyto check that a recipe's adapter attaches before you download weights. - Configuration: every key a recipe can set.
- Quant Menu: what each
quantizationvalue does. - Config refusals: the settings v0.77.0 refuses at config load; run
soup doctor --config soup.yamlon a recipe you have edited.
Soup is free and Apache-2.0. If it saved you a training run, starring the repo costs nothing and helps most. You can also fund the GPU time behind the work a 4 GB laptop cannot reach.