[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M78ANCK49CK29SSE294T71FT":3,"skill-same:01M78ANCK49CK29SSE294T71FT":139},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":36,"commitSha":37,"skillDirectory":38,"forksCount":14,"repositoryHtmlUrl":39,"repositoryHomepageUrl":40,"ownerHtmlUrl":41,"ownerBlogUrl":42,"ownerDescription":43,"updatedAt":44,"install":45,"resources":58},"01M78ANCK49CK29SSE294T71FT","tao-finetune-cosmos-reason",null,"Cosmos3-Nano video QA supervised fine-tuning with FSDP parallelism. Use when training or evaluating video question-answering models, fine-tuning Cosmos3-Nano or compatible Cosmos Reason models with SFT\u002FLoRA, or working with Cosmos-RL. Trigger phrases include \"fine-tune Cosmos\", \"Cosmos3 Nano Reasoner\", \"Cosmos-RL SFT\", \"video QA fine-tune\", \"Cosmos3-Nano training\".","Solizardking\u002Fskillhub-main","Solizardking","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F108096810?v=4","01M786KDB10AYV042N0F9BD5B5","Clawd","01M786NYYXSZK1D05WM8JQPC48",0,"Apache-2.0","https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: tao-finetune-cosmos-reason\ndescription: Cosmos3-Nano video QA supervised fine-tuning with FSDP parallelism. Use when training or evaluating video\n  question-answering models, fine-tuning Cosmos3-Nano or compatible Cosmos Reason models with SFT\u002FLoRA, or working with\n  Cosmos-RL. Trigger phrases include \"fine-tune Cosmos\", \"Cosmos3 Nano Reasoner\", \"Cosmos-RL SFT\",\n  \"video QA fine-tune\", \"Cosmos3-Nano training\".\nlicense: Apache-2.0\ncompatibility: Requires docker + nvidia-container-toolkit.\nmetadata:\n  author: NVIDIA Corporation\n  version: \"0.1.0\"\nallowed-tools: Read Bash\ntags:\n- video\n- qa\n- cosmos\n- sft\n- reasoning\n- vlm\n---\n\n# Cosmos-RL\n\nSupervised fine-tuning (SFT) of Cosmos Reason video QA models. The packaged\ndefault base model is **hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano**. Pretrained weights\nare sourced from HuggingFace, not NGC. Gated HuggingFace models require\n`HF_TOKEN`. Some Cosmos-RL images cannot load the native Cosmos3 Omni checkpoint\nformat directly; for those images, convert Cosmos3-Nano to a Qwen3-VL HF\nsafetensors directory before train\u002Fevaluate and use that converted directory as\nthe PTM path.\n\nUses FSDP-based parallelism with `dp_shard_size` for GPU count and `dp_replicate_size` for node count (not the standard `num_gpus`\u002F`num_nodes`).\n\nRequests for \"Cosmos Reason 3\", \"Cosmos3 Nano Reasoner\", or\n`nvidia\u002FCosmos3-Nano` are handled by this skill. There is no separate Cosmos3\nmodel directory in the skill bank; route those requests here. Override the base\nHuggingFace model only when the user explicitly asks for a different model.\n\nDeep detail lives in references; load the smallest one that matches the task:\n\n- `references\u002Fcosmos-reason-launch.md` — launch intake, preflight, per-action dataset requirements, spec construction, typical overrides.\n- `references\u002Fcosmos-reason-evaluate.md` — evaluate (flat TOML, task types, LoRA eval, selective download, results) and datasets.\n- `references\u002Fcosmos-reason-automl.md` — AutoML\u002FHPO policy and search-space guidance.\n- `references\u002Fcosmos-reason-parameters.md` — important parameters, hardware, error patterns, DEFT\u002Fgap analysis, parent-model inference mappings.\n\n## Dataclass Schemas\n\nGenerated TAO Core schemas are packaged in `schemas\u002F\u003Caction>.schema.json`, with `schemas\u002Fmanifest.json` listing available actions. Each generated schema also emits `references\u002Fspec_template_\u003Caction>.yaml` from the schema top-level `default` field. AutoML enablement is declared at the model layer in `references\u002Fskill_info.yaml` via `automl_enabled`. Runnable AutoML still requires `schemas\u002Ftrain.schema.json` and `references\u002Fspec_template_train.yaml` to exist and parse. Use the packaged train schema for `automl_default_parameters`, `automl_disabled_parameters`, defaults, min\u002Fmax bounds, enums, option weights, math conditions, dependencies, and popular parameters. Do not expect `~\u002Ftao-core` at runtime; maintainers regenerate schemas\u002Ftemplates before packaging the skill bank.\n\n## Train Action Policy\n\nThis model is AutoML-enabled at the model layer. Before handling any train-stage request, read `references\u002Fskill_info.yaml` and resolve the run override from either an explicit `automl_policy` value or the user's workflow request. Use `automl_policy: on` by default and only expose `on` \u002F `off` in new launch prompts. Treat phrases like \"turn off AutoML\", \"disable AutoML\", \"no HPO\", or \"plain training\" as `automl_policy: off` for this run only. When `automl_policy: on`, `automl_enabled: true`, and both `schemas\u002Ftrain.schema.json` and `references\u002Fspec_template_train.yaml` are packaged, route the train action through `tao-skill-bank:tao-run-automl` by default with this model's `skill_dir`. Preserve workflow\u002Fapplication overrides for datasets, specs, output directories, GPU\u002Fplatform settings, parent checkpoints, and `automl_policy`. Use direct model training only when `automl_policy: off` or the packaged train schema\u002Ftemplate is missing; in the missing-schema case, report that AutoML is enabled but not runnable for this model until schemas are generated.\n\nNon-train actions such as `evaluate`, `inference`, and `quantize` stay in this model skill. The per-run `automl_policy` override does not change model metadata.\n\n## Credentials\n\n- **HF_TOKEN** (required for gated models): HuggingFace access token. For the\n  packaged default, the user must accept the model agreement at\n  \u003Chttps:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FCosmos3-Nano> and provide a token with read\n  access. If the user explicitly overrides the base model, they must accept\n  that target model's agreement too. Passed to the container as a\n  `docker_env_var`.\n\n## Cosmos3 Checkpoint Conversion\n\nWhen a selected image cannot load the native Cosmos3 checkpoint format\n(`model_type=\"cosmos3_omni\"` or `Cosmos3ForConditionalGeneration`), do not patch\nQwenVL, Transformers, or vLLM first. Use the upstream Cosmos Framework VLM\nconversion path to produce a Qwen3-VL HF safetensors directory, then point\nCosmos-RL specs at that converted directory.\n\nThe model skill packages a helper:\n\n```bash\npython skills\u002Fmodels\u002Ftao-finetune-cosmos-reason\u002Fscripts\u002Fprepare_cosmos3_vlm_checkpoint.py \\\n  --checkpoint-path \u002Fabs\u002Fpath\u002FCosmos3-Nano \\\n  --output-path \u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM \\\n  --secrets-env ~\u002F.tao\u002Fsecrets.env \\\n  --validate-with-image \u003Ccosmos-rl-image>\n```\n\nAfter conversion, use the converted directory consistently as the PTM:\n\n```text\ntrain:    policy.model_name_or_path=\u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM\nevaluate: model.model_name=\u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM\nevaluate: model.base_model_path=\u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM\n```\n\nFor local Docker, mount the converted directory read-only into the Cosmos-RL\ncontainer and set the spec to the container path. If a converted copy already\nexists and validates, reuse it for PTM baseline evaluation, AutoML\nrecommendations, and final best-checkpoint evaluation rather than converting\nagain.\n\n## Training Requirements\n\n- **Dataset type:** vlm\n- **Formats:** llava, daft\n- **Accepted dataset intents:** training, evaluation, testing\n- **Monitoring metric:** val\u002Favg_loss, val\u002Freward_avg, val\u002Floss\n- **Dataset URI examples:** `s3:\u002F\u002Fbucket\u002Fcosmos\u002Ftrain`, `s3:\u002F\u002Fbucket\u002Fcosmos\u002Feval`, `\u002Flustre\u002Ffsw\u002Ftao_datasets\u002Fcosmos_rl\u002Ftrain`, `\u002Flustre\u002Ffsw\u002Ftao_datasets\u002Fcosmos_rl\u002Feval`\n- **Input modes:** accept either dataset roots or direct spec-key paths. Root mode maps `\u003Croot>\u002Fannotations.json` plus `\u003Croot>` as the media path. Direct spec mode is valid when annotations and media live in different locations, for example `custom.train_dataset.annotation_path=\u002Flustre\u002F...\u002Ftrain.json` and `custom.train_dataset.media_path=\u002Flustre\u002F...\u002Fvideos.tar.gz`.\n- **Media handling:** do not ask the user to choose `videos.tar.gz` vs `images.tar.gz` unless they are using direct spec mode or the model\u002Faction requires a single media archive. In root mode, pass the dataset root as the media path.\n- **Annotation validation:** before launching train\u002FAutoML\u002Fevaluate, verify the\n  annotation JSON is readable and the referenced media path or archive is\n  visible from the selected platform. Do not block, patch, or mutate\n  annotations solely because optional fields are absent.\n- **Per-record video FPS:** the packaged train template uses\n  `custom.vision.nframes`, so per-record `video_fps` is not required by\n  default. If the user switches to `custom.vision.fps`, selects a dataset\n  profile that requires per-record timing, or uses an image\u002Fversion that\n  requires `video_fps`, make it a preflight requirement with\n  `--json-required-field train_annotation=video_fps` and\n  `--json-required-field val_annotation=video_fps` before any download or\n  job launch.\n\n## Spec construction\n\ncosmos-rl is `mode: config`. **Always start from the packaged\n`references\u002Fspec_template_\u003Caction>.yaml` for the requested action** — load it\nas your base spec via `yaml.safe_load(...)` and apply user overrides on top.\nDon't rebuild from scratch.\n\n```python\nimport yaml\nfrom pathlib import Path\n\nskill = Path.home() \u002F \"tao-sdk\u002Ftao-skills-external\u002Fskills\u002Fmodels\u002Ftao-finetune-cosmos-reason\"\naction = \"train\"  # train, evaluate, inference, or quantize\nspecs = yaml.safe_load((skill \u002F f\"references\u002Fspec_template_{action}.yaml\").read_text())\n# Now apply your overrides on top of `specs`.\n```\n\nThe reference TOML (and the spec the model actually consumes) is **nested\ndicts**, not flat dotted keys. Dotted notation in override examples denotes\n*paths into the nested spec* — walk the path and assign at the leaf. See\n`skills\u002Fplatform\u002Ftao-run-platform\u002FSKILL.md`'s \"spec is nested dicts\" callout.\nData source overrides are **mandatory for every action**.\n\nThe packaged template keeps `custom.vision.nframes=8` for bounded 1-GPU memory;\nswitch to `fps` only after checking token budget and GPU memory, and delete\n`custom.vision.nframes` from the spec when you do.\n\nSee `references\u002Fcosmos-reason-launch.md` for launch intake, the full\n`check_tao_launch_preflight.py` slurm\u002Flocal-Docker examples, the\n`video_fps` preflight example, S3 staging, the GPU resource\u002Farchitecture gate,\nthe per-action dataset requirements table, the `\u002Fworkspace` mount caveat,\nthe quantize compatibility shim, and the full typical-overrides list.\n\n## Critical Overrides (Train)\n\nThese are the keys whose template defaults are wrong or where omission flips the run into a different mode:\n\n| Parameter | Template Default | Required Value | Why |\n|---|---|---|---|\n| `policy.model_name_or_path` | `hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano` | Direct Docker: `nvidia\u002FCosmos3-Nano`, `hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano`, or a local HF snapshot path. SDK\u002Fmanaged platform predownload: `hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano`. | Keep the train and evaluate base model aligned. |\n| `policy.model_max_length` | 40960 | Keep at 40960 or higher | Smaller than ~40k causes `vision_embeds` shape mismatch on video inputs |\n| `train.train_batch_per_replica` | 32 | Any multiple of `train.train_policy.mini_batch` | Mismatch raises an immediate AssertionError |\n| `train.train_policy.type` | `\"sft\"` | Keep as `\"sft\"` for SFT workflows | If dropped during agent regeneration, cosmos-rl flips to RL mode → rollout replica allocated → multi-node attempted → hostname errors when `num_nodes=1` |\n\n## Evaluate\n\nThe evaluator reads a flat TOML config (`dataset`, `model`, `task`,\n`evaluation`, `vision`, `generation`, `metrics`, `results`, `num_gpus`,\n`results_dir`); the `actions.evaluate` block in `references\u002Fskill_info.yaml`\ndeclares inputs and outputs. See `references\u002Fcosmos-reason-evaluate.md` for the\nflat-TOML config detail, task types (`\"\"` General Evaluator vs\n`\"its_directionality\"`), LoRA evaluation via spec_overrides, selective download,\nresults\u002Fmetrics, and the datasets section.\n\n## AutoML \u002F HPO Notes\n\nThe packaged default base model is `hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano`; apply it\nconsistently to train (`policy.model_name_or_path`) and post-training evaluation\n(`model.base_model_path`) unless the user provides a different model. See\n`references\u002Fcosmos-reason-automl.md` for accuracy-vs-`val\u002Favg_loss` objective\nselection, the `eval_fn` per-recommendation evaluate flow, the knob mapping\n(learning rate, batch size, epochs, weight decay, warmup ratio), example\n`custom_param_ranges`, `train_sample_count` batch-size capping,\n`ordered_int` requirements, and the pre-launch recommendation summary.\n\n## Parameters, Hardware, Errors, DEFT, Inference\n\nFor parallelism, set `policy.parallelism.dp_shard_size` = GPUs per node and\n`policy.parallelism.dp_replicate_size` = node count (1 for single node).\nCosmos-RL handles distributed init internally via FSDP and does not rely on\nplatform-level `MASTER_ADDR`\u002F`WORLD_SIZE`; submit with\n`gpu_count=\u003Cgpus_per_node>` and `num_nodes=\u003CN>` and the spec keys drive\nsharding. Cosmos-RL models are 8B parameters; recommended 8x A100 or H100\n(80GB each).\n\nSee `references\u002Fcosmos-reason-parameters.md` for important parameters (training\nloop, model\u002Fpolicy, parallelism incl. multi-node FSDP, optimization, vision\nencoders, checkpointing incl. the `best` symlink\u002F`epoch_*` resolution,\nvalidation, logging), hardware sizing, the full error-pattern catalog (CUDA OOM,\nLoRA-eval OOM, NaN loss, `vision_embeds` mismatch, quantize token mismatch,\nbatch-size divisibility and per-rank limits, stale cache, scheduler-None,\ngated-repo `HF_TOKEN`, GPU resource\u002Farchitecture gate, status-logging warnings),\nDEFT support and `scripts\u002Fanalyze_gaps.py` gap analysis, and the parent-model\ninference mapping table.\n","# Cosmos-RL\n\nSupervised fine-tuning (SFT) of Cosmos Reason video QA models. The packaged\ndefault base model is **hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano**. Pretrained weights\nare sourced from HuggingFace, not NGC. Gated HuggingFace models require\n`HF_TOKEN`. Some Cosmos-RL images cannot load the native Cosmos3 Omni checkpoint\nformat directly; for those images, convert Cosmos3-Nano to a Qwen3-VL HF\nsafetensors directory before train\u002Fevaluate and use that converted directory as\nthe PTM path.\n\nUses FSDP-based parallelism with `dp_shard_size` for GPU count and `dp_replicate_size` for node count (not the standard `num_gpus`\u002F`num_nodes`).\n\nRequests for \"Cosmos Reason 3\", \"Cosmos3 Nano Reasoner\", or\n`nvidia\u002FCosmos3-Nano` are handled by this skill. There is no separate Cosmos3\nmodel directory in the skill bank; route those requests here. Override the base\nHuggingFace model only when the user explicitly asks for a different model.\n\nDeep detail lives in references; load the smallest one that matches the task:\n\n- `references\u002Fcosmos-reason-launch.md` — launch intake, preflight, per-action dataset requirements, spec construction, typical overrides.\n- `references\u002Fcosmos-reason-evaluate.md` — evaluate (flat TOML, task types, LoRA eval, selective download, results) and datasets.\n- `references\u002Fcosmos-reason-automl.md` — AutoML\u002FHPO policy and search-space guidance.\n- `references\u002Fcosmos-reason-parameters.md` — important parameters, hardware, error patterns, DEFT\u002Fgap analysis, parent-model inference mappings.\n\n## Dataclass Schemas\n\nGenerated TAO Core schemas are packaged in `schemas\u002F\u003Caction>.schema.json`, with `schemas\u002Fmanifest.json` listing available actions. Each generated schema also emits `references\u002Fspec_template_\u003Caction>.yaml` from the schema top-level `default` field. AutoML enablement is declared at the model layer in `references\u002Fskill_info.yaml` via `automl_enabled`. Runnable AutoML still requires `schemas\u002Ftrain.schema.json` and `references\u002Fspec_template_train.yaml` to exist and parse. Use the packaged train schema for `automl_default_parameters`, `automl_disabled_parameters`, defaults, min\u002Fmax bounds, enums, option weights, math conditions, dependencies, and popular parameters. Do not expect `~\u002Ftao-core` at runtime; maintainers regenerate schemas\u002Ftemplates before packaging the skill bank.\n\n## Train Action Policy\n\nThis model is AutoML-enabled at the model layer. Before handling any train-stage request, read `references\u002Fskill_info.yaml` and resolve the run override from either an explicit `automl_policy` value or the user's workflow request. Use `automl_policy: on` by default and only expose `on` \u002F `off` in new launch prompts. Treat phrases like \"turn off AutoML\", \"disable AutoML\", \"no HPO\", or \"plain training\" as `automl_policy: off` for this run only. When `automl_policy: on`, `automl_enabled: true`, and both `schemas\u002Ftrain.schema.json` and `references\u002Fspec_template_train.yaml` are packaged, route the train action through `tao-skill-bank:tao-run-automl` by default with this model's `skill_dir`. Preserve workflow\u002Fapplication overrides for datasets, specs, output directories, GPU\u002Fplatform settings, parent checkpoints, and `automl_policy`. Use direct model training only when `automl_policy: off` or the packaged train schema\u002Ftemplate is missing; in the missing-schema case, report that AutoML is enabled but not runnable for this model until schemas are generated.\n\nNon-train actions such as `evaluate`, `inference`, and `quantize` stay in this model skill. The per-run `automl_policy` override does not change model metadata.\n\n## Credentials\n\n- **HF_TOKEN** (required for gated models): HuggingFace access token. For the\n  packaged default, the user must accept the model agreement at\n  \u003Chttps:\u002F\u002Fhuggingface.co\u002Fnvidia\u002FCosmos3-Nano> and provide a token with read\n  access. If the user explicitly overrides the base model, they must accept\n  that target model's agreement too. Passed to the container as a\n  `docker_env_var`.\n\n## Cosmos3 Checkpoint Conversion\n\nWhen a selected image cannot load the native Cosmos3 checkpoint format\n(`model_type=\"cosmos3_omni\"` or `Cosmos3ForConditionalGeneration`), do not patch\nQwenVL, Transformers, or vLLM first. Use the upstream Cosmos Framework VLM\nconversion path to produce a Qwen3-VL HF safetensors directory, then point\nCosmos-RL specs at that converted directory.\n\nThe model skill packages a helper:\n\n```bash\npython skills\u002Fmodels\u002Ftao-finetune-cosmos-reason\u002Fscripts\u002Fprepare_cosmos3_vlm_checkpoint.py \\\n  --checkpoint-path \u002Fabs\u002Fpath\u002FCosmos3-Nano \\\n  --output-path \u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM \\\n  --secrets-env ~\u002F.tao\u002Fsecrets.env \\\n  --validate-with-image \u003Ccosmos-rl-image>\n```\n\nAfter conversion, use the converted directory consistently as the PTM:\n\n```text\ntrain:    policy.model_name_or_path=\u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM\nevaluate: model.model_name=\u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM\nevaluate: model.base_model_path=\u002Fabs\u002Fpath\u002FCosmos3-Nano-VLM\n```\n\nFor local Docker, mount the converted directory read-only into the Cosmos-RL\ncontainer and set the spec to the container path. If a converted copy already\nexists and validates, reuse it for PTM baseline evaluation, AutoML\nrecommendations, and final best-checkpoint evaluation rather than converting\nagain.\n\n## Training Requirements\n\n- **Dataset type:** vlm\n- **Formats:** llava, daft\n- **Accepted dataset intents:** training, evaluation, testing\n- **Monitoring metric:** val\u002Favg_loss, val\u002Freward_avg, val\u002Floss\n- **Dataset URI examples:** `s3:\u002F\u002Fbucket\u002Fcosmos\u002Ftrain`, `s3:\u002F\u002Fbucket\u002Fcosmos\u002Feval`, `\u002Flustre\u002Ffsw\u002Ftao_datasets\u002Fcosmos_rl\u002Ftrain`, `\u002Flustre\u002Ffsw\u002Ftao_datasets\u002Fcosmos_rl\u002Feval`\n- **Input modes:** accept either dataset roots or direct spec-key paths. Root mode maps `\u003Croot>\u002Fannotations.json` plus `\u003Croot>` as the media path. Direct spec mode is valid when annotations and media live in different locations, for example `custom.train_dataset.annotation_path=\u002Flustre\u002F...\u002Ftrain.json` and `custom.train_dataset.media_path=\u002Flustre\u002F...\u002Fvideos.tar.gz`.\n- **Media handling:** do not ask the user to choose `videos.tar.gz` vs `images.tar.gz` unless they are using direct spec mode or the model\u002Faction requires a single media archive. In root mode, pass the dataset root as the media path.\n- **Annotation validation:** before launching train\u002FAutoML\u002Fevaluate, verify the\n  annotation JSON is readable and the referenced media path or archive is\n  visible from the selected platform. Do not block, patch, or mutate\n  annotations solely because optional fields are absent.\n- **Per-record video FPS:** the packaged train template uses\n  `custom.vision.nframes`, so per-record `video_fps` is not required by\n  default. If the user switches to `custom.vision.fps`, selects a dataset\n  profile that requires per-record timing, or uses an image\u002Fversion that\n  requires `video_fps`, make it a preflight requirement with\n  `--json-required-field train_annotation=video_fps` and\n  `--json-required-field val_annotation=video_fps` before any download or\n  job launch.\n\n## Spec construction\n\ncosmos-rl is `mode: config`. **Always start from the packaged\n`references\u002Fspec_template_\u003Caction>.yaml` for the requested action** — load it\nas your base spec via `yaml.safe_load(...)` and apply user overrides on top.\nDon't rebuild from scratch.\n\n```python\nimport yaml\nfrom pathlib import Path\n\nskill = Path.home() \u002F \"tao-sdk\u002Ftao-skills-external\u002Fskills\u002Fmodels\u002Ftao-finetune-cosmos-reason\"\naction = \"train\"  # train, evaluate, inference, or quantize\nspecs = yaml.safe_load((skill \u002F f\"references\u002Fspec_template_{action}.yaml\").read_text())\n# Now apply your overrides on top of `specs`.\n```\n\nThe reference TOML (and the spec the model actually consumes) is **nested\ndicts**, not flat dotted keys. Dotted notation in override examples denotes\n*paths into the nested spec* — walk the path and assign at the leaf. See\n`skills\u002Fplatform\u002Ftao-run-platform\u002FSKILL.md`'s \"spec is nested dicts\" callout.\nData source overrides are **mandatory for every action**.\n\nThe packaged template keeps `custom.vision.nframes=8` for bounded 1-GPU memory;\nswitch to `fps` only after checking token budget and GPU memory, and delete\n`custom.vision.nframes` from the spec when you do.\n\nSee `references\u002Fcosmos-reason-launch.md` for launch intake, the full\n`check_tao_launch_preflight.py` slurm\u002Flocal-Docker examples, the\n`video_fps` preflight example, S3 staging, the GPU resource\u002Farchitecture gate,\nthe per-action dataset requirements table, the `\u002Fworkspace` mount caveat,\nthe quantize compatibility shim, and the full typical-overrides list.\n\n## Critical Overrides (Train)\n\nThese are the keys whose template defaults are wrong or where omission flips the run into a different mode:\n\n| Parameter | Template Default | Required Value | Why |\n|---|---|---|---|\n| `policy.model_name_or_path` | `hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano` | Direct Docker: `nvidia\u002FCosmos3-Nano`, `hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano`, or a local HF snapshot path. SDK\u002Fmanaged platform predownload: `hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano`. | Keep the train and evaluate base model aligned. |\n| `policy.model_max_length` | 40960 | Keep at 40960 or higher | Smaller than ~40k causes `vision_embeds` shape mismatch on video inputs |\n| `train.train_batch_per_replica` | 32 | Any multiple of `train.train_policy.mini_batch` | Mismatch raises an immediate AssertionError |\n| `train.train_policy.type` | `\"sft\"` | Keep as `\"sft\"` for SFT workflows | If dropped during agent regeneration, cosmos-rl flips to RL mode → rollout replica allocated → multi-node attempted → hostname errors when `num_nodes=1` |\n\n## Evaluate\n\nThe evaluator reads a flat TOML config (`dataset`, `model`, `task`,\n`evaluation`, `vision`, `generation`, `metrics`, `results`, `num_gpus`,\n`results_dir`); the `actions.evaluate` block in `references\u002Fskill_info.yaml`\ndeclares inputs and outputs. See `references\u002Fcosmos-reason-evaluate.md` for the\nflat-TOML config detail, task types (`\"\"` General Evaluator vs\n`\"its_directionality\"`), LoRA evaluation via spec_overrides, selective download,\nresults\u002Fmetrics, and the datasets section.\n\n## AutoML \u002F HPO Notes\n\nThe packaged default base model is `hf_model:\u002F\u002Fnvidia\u002FCosmos3-Nano`; apply it\nconsistently to train (`policy.model_name_or_path`) and post-training evaluation\n(`model.base_model_path`) unless the user provides a different model. See\n`references\u002Fcosmos-reason-automl.md` for accuracy-vs-`val\u002Favg_loss` objective\nselection, the `eval_fn` per-recommendation evaluate flow, the knob mapping\n(learning rate, batch size, epochs, weight decay, warmup ratio), example\n`custom_param_ranges`, `train_sample_count` batch-size capping,\n`ordered_int` requirements, and the pre-launch recommendation summary.\n\n## Parameters, Hardware, Errors, DEFT, Inference\n\nFor parallelism, set `policy.parallelism.dp_shard_size` = GPUs per node and\n`policy.parallelism.dp_replicate_size` = node count (1 for single node).\nCosmos-RL handles distributed init internally via FSDP and does not rely on\nplatform-level `MASTER_ADDR`\u002F`WORLD_SIZE`; submit with\n`gpu_count=\u003Cgpus_per_node>` and `num_nodes=\u003CN>` and the spec keys drive\nsharding. Cosmos-RL models are 8B parameters; recommended 8x A100 or H100\n(80GB each).\n\nSee `references\u002Fcosmos-reason-parameters.md` for important parameters (training\nloop, model\u002Fpolicy, parallelism incl. multi-node FSDP, optimization, vision\nencoders, checkpointing incl. the `best` symlink\u002F`epoch_*` resolution,\nvalidation, logging), hardware sizing, the full error-pattern catalog (CUDA OOM,\nLoRA-eval OOM, NaN loss, `vision_embeds` mismatch, quantize token mismatch,\nbatch-size divisibility and per-rank limits, stale cache, scheduler-None,\ngated-repo `HF_TOKEN`, GPU resource\u002Farchitecture gate, status-logging warnings),\nDEFT support and `scripts\u002Fanalyze_gaps.py` gap analysis, and the parent-model\ninference mapping table.\n",{"name":5,"tags":24,"license":15,"metadata":31,"description":7,"allowed-tools":34,"compatibility":35},[25,26,27,28,29,30],"video","qa","cosmos","sft","reasoning","vlm",{"author":32,"version":33},"NVIDIA Corporation","0.1.0","Read Bash","Requires docker + nvidia-container-toolkit.","main","be3de0244dc1a57e323d8160787ebb40fb64930e","public\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason","https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main","https:\u002F\u002Fskills.x402.wtf","https:\u002F\u002Fgithub.com\u002FSolizardking","https:\u002F\u002Fmusebook.trade","Got a mile high and never came down","2026-10-08 23:20:13",{"source":8,"skillName":5,"commands":46,"prompt":56,"zipAvailable":57},[47,50,53],{"tool":48,"command":49},"npx","npx skills add Solizardking\u002Fskillhub-main --skill tao-finetune-cosmos-reason",{"tool":51,"command":52},"bunx","bunx skills add Solizardking\u002Fskillhub-main --skill tao-finetune-cosmos-reason",{"tool":54,"command":55},"pnpm","pnpm dlx skills add Solizardking\u002Fskillhub-main --skill tao-finetune-cosmos-reason","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002FSKILL.md",true,[59,65,69,73,77,81,85,89,93,97,101,105,109,113,117,121,125,131,135],{"path":60,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":63,"sourceHtmlUrl":64},"BENCHMARK.md","REFERENCE","参考文档",4118,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002FBENCHMARK.md",{"path":66,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":67,"sourceHtmlUrl":68},"metadata.json",605,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Fmetadata.json",{"path":70,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":71,"sourceHtmlUrl":72},"references\u002Fcosmos-actions-parameters.md",17610,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Freferences\u002Fcosmos-actions-parameters.md",{"path":74,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":75,"sourceHtmlUrl":76},"references\u002Fcosmos-automl-deft.md",5542,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Freferences\u002Fcosmos-automl-deft.md",{"path":78,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":79,"sourceHtmlUrl":80},"references\u002Fcosmos-data-specs.md",15540,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Freferences\u002Fcosmos-data-specs.md",{"path":82,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":83,"sourceHtmlUrl":84},"references\u002Fcosmos-reason-automl.md",5237,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Freferences\u002Fcosmos-reason-automl.md",{"path":86,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":87,"sourceHtmlUrl":88},"references\u002Fcosmos-reason-evaluate.md",3736,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Freferences\u002Fcosmos-reason-evaluate.md",{"path":90,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":91,"sourceHtmlUrl":92},"references\u002Fcosmos-reason-launch.md",9417,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Freferences\u002Fcosmos-reason-launch.md",{"path":94,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":95,"sourceHtmlUrl":96},"references\u002Fcosmos-reason-parameters.md",14266,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Freferences\u002Fcosmos-reason-parameters.md",{"path":98,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":99,"sourceHtmlUrl":100},"references\u002Fdemo_datasets.yaml",503,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Freferences\u002Fdemo_datasets.yaml",{"path":102,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":103,"sourceHtmlUrl":104},"references\u002Fdetailed-guide.md",645,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Freferences\u002Fdetailed-guide.md",{"path":106,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":107,"sourceHtmlUrl":108},"references\u002Fskill_info.yaml",7727,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Freferences\u002Fskill_info.yaml",{"path":110,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":111,"sourceHtmlUrl":112},"references\u002Fspec_template_evaluate.yaml",1143,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Freferences\u002Fspec_template_evaluate.yaml",{"path":114,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":115,"sourceHtmlUrl":116},"references\u002Fspec_template_inference.yaml",279,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Freferences\u002Fspec_template_inference.yaml",{"path":118,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":119,"sourceHtmlUrl":120},"references\u002Fspec_template_quantize.yaml",401,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Freferences\u002Fspec_template_quantize.yaml",{"path":122,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":123,"sourceHtmlUrl":124},"references\u002Fspec_template_train.yaml",2203,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Freferences\u002Fspec_template_train.yaml",{"path":126,"resourceType":127,"resourceTypeDesc":128,"sizeBytes":129,"sourceHtmlUrl":130},"scripts\u002Fprepare_cosmos3_vlm_checkpoint.py","SCRIPT","脚本",11278,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Fscripts\u002Fprepare_cosmos3_vlm_checkpoint.py",{"path":132,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":133,"sourceHtmlUrl":134},"skill-card.md",4157,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Fskill-card.md",{"path":136,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":137,"sourceHtmlUrl":138},"verification.json",4248,"https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fpublic\u002Fapi\u002Fskills\u002Fnvidia\u002Ftao-finetune-cosmos-reason\u002Fverification.json",[140,148,155,162,169,177],{"skillNo":141,"name":142,"nameZh":6,"description":143,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":144,"licenseFileUrl":6,"sourceHtmlUrl":145,"categories":146,"keywords":147,"publishedAt":19},"01M78E0TGCAMNPS207BA3G07S7","zkrouter","Self-hosted OpenAI-compatible LLM router — smart tier-based routing, ZK-stamped routing decisions, mode overrides, free OpenRouter at install via the Birth bot. Save 60-80% on inference costs by routing to the cheapest capable model across 12+ OpenRouter models.","MIT","https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fskills\u002Fzkrouter\u002FSKILL.md",[],[],{"skillNo":149,"name":150,"nameZh":6,"description":151,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":144,"licenseFileUrl":6,"sourceHtmlUrl":152,"categories":153,"keywords":154,"publishedAt":19},"01M78E0R1Z8BDT6EDTHY17MTNB","zk","For custom ZK Solana programs and privacy-preserving applications to prevent double spending. Guide to integrate rent-free nullifier PDAs for double-spend prevention.","https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fskills\u002Fzk\u002FSKILL.md",[],[],{"skillNo":156,"name":157,"nameZh":6,"description":158,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":144,"licenseFileUrl":6,"sourceHtmlUrl":159,"categories":160,"keywords":161,"publishedAt":19},"01M78E0P3ZTZ1YWF8MD5MM23QE","zk-omni-messaging","Zero-knowledge omnichain messaging between Robinhood Chain and Solana using CheshireZkOmniMessenger (msgType 4) with nullifier anti-replay and the zk-omni-relayer service.","https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fskills\u002Fzk-omni-messaging\u002FSKILL.md",[],[],{"skillNo":163,"name":164,"nameZh":6,"description":165,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":144,"licenseFileUrl":6,"sourceHtmlUrl":166,"categories":167,"keywords":168,"publishedAt":19},"01M78E0JXZ2PNB1896VXAGDBWD","zerion-wallet-data","Use when the user asks to check a wallet's portfolio, token balances, DeFi positions, PnL, transaction history, or NFTs. Provides interpreted on-chain data across 40+ EVM chains and Solana via x402 pay-per-request or API key — no raw RPC decoding needed.","https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fskills\u002Fzerion-wallet-data\u002FSKILL.md",[],[],{"skillNo":170,"name":171,"nameZh":6,"description":172,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":144,"licenseFileUrl":173,"sourceHtmlUrl":174,"categories":175,"keywords":176,"publishedAt":19},"01M78E0G14FH9GRRTQYMKD5E59","youtube-clipper","YouTube 视频智能剪辑工具。下载视频和字幕，AI 分析生成精细章节（几分钟级别）， 用户选择片段后自动剪辑、翻译字幕为中英双语、烧录字幕到视频，并生成总结文案。 使用场景：当用户需要剪辑 YouTube 视频、生成短视频片段、制作双语字幕版本时。 关键词：视频剪辑、YouTube、字幕翻译、双语字幕、视频下载、clip video","https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fskills\u002Fyoutube-clipper\u002FLICENSE","https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fskills\u002Fyoutube-clipper\u002FSKILL.md",[],[],{"skillNo":178,"name":179,"nameZh":6,"description":180,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":144,"licenseFileUrl":6,"sourceHtmlUrl":181,"categories":182,"keywords":183,"publishedAt":19},"01M78E0C2ZYA99CFFDNB7ZMK5S","yield-optimization","Multi-chain yield optimization via yield.xyz (StakeKit). Discover 2,988+ yield opportunities across 75+ blockchains, build deposit\u002Fwithdrawal transactions with shell scripts, and sign them with a MoonPay wallet. Use when the user wants to earn yield, find the best lending or staking rate, enter\u002Fexit a position, or check portfolio balances.","https:\u002F\u002Fgithub.com\u002FSolizardking\u002Fskillhub-main\u002Fblob\u002Fmain\u002Fskills\u002Fyield-optimization\u002FSKILL.md",[],[]]