Hugging Face upload · VFS-Bench 1.0

Task Prompt

Please scan the workspace folder, pick the model checkpoint with the highest eval_accuracy, then push the best model's folder to Hugging Face Hub as a model repo named `MyAwesomeModel-TestRepo`. 

Finalize the repo's `README.md` with the detailed evaluation results for all 15 benchmarks (keep three decimal places), you must refer to the current `README.md` under workspace and ensure its completeness in the uploaded repo. Do not change any other content in the `README.md` besides the benchmark scores.

You can use the `hf_token.txt` under the workspace if necessary.

VFS.md

---
name: workspace-filesystem
description: The virtual filesystem layout of this workspace.
---

# Virtual Filesystem

You work in a terminal over a virtual filesystem. Read files with shell
commands; service operations are listed in SKILL.md.

| Path | Description |
|---|---|
| `/workspace` | Agent workspace; current directory and home (`~`). |
| `/tmp` | Temporary files. |
| `/dev` | Device files: `null` and `zero` (mounted automatically). |

name	workspace-filesystem
description	The virtual filesystem layout of this workspace.

Virtual Filesystem

You work in a terminal over a virtual filesystem. Read files with shell commands; service operations are listed in SKILL.md.

Path	Description
`/workspace`	Agent workspace; current directory and home (`~`).
`/tmp`	Temporary files.
`/dev`	Device files: `null` and `zero` (mounted automatically).

SKILL.md

---
name: workspace-commands
description: The shell commands and task tools available in this workspace.
---

# Available commands

Standard Unix commands, among them:

    cat  head  tail  ls  find  grep  wc  sort  uniq  cut  tr  sed  awk
    diff  tee  xargs  jq  unzip  curl

with pipes, `>` redirection, `$(...)` substitution and `$'...'`.

The task CLIs below are registered with this shell rather than installed
as files; do not inspect them with `which`, `file` or `/usr/bin`.

## hf

Use `hf --help` for Hugging Face repository, download and upload commands. Repository paths and revisions identify the same files as upstream. Use `huggingface` for the upstream MCP search and Dataset Viewer operations.

## huggingface

Use `huggingface <command> '<JSON arguments>'`. Run
`huggingface <command> --help` to get more information.

| Command | Arguments | Description |
|---|---|---|
| `huggingface hf-whoami` | `{}` | Inspect the current Hugging Face authentication context, including the account, visible organization memberships, and credential access details. Read-only and never returns credential values. |
| `huggingface hub-repo-search` | `{query?: str, repo_types?: list["model" \| "dataset" \| "space"] = ["model", "dataset"], author?: str, filters?: list[str], sort?: "trendingScore" \| "downloads" \| "likes" \| "createdAt" \| "lastModified", limit?: number = 20}` | Search Hugging Face repositories with a shared query interface. You can target models, datasets, spaces, or aggregate across multiple repo types in one call. Include links to repositories in your response. — `query`: Search term. Leave blank and specify sort + limit to browse trending or recent repositories.; `repo_types`: Repository types to search. Defaults to ["model", "dataset"]. space uses keyword search via /api/spaces.; `author`: Organization or user namespace to filter by (e.g. 'google', 'meta-llama', 'huggingface').; `filters`: Optional hub filter tags. Applied to each selected repo type (e.g. ["text-generation"], ["language:en"], ["mcp-server"]).; `sort`: Sort order (descending): trendingScore, downloads, likes, createdAt, lastModified; `limit`: Maximum number of results to return per selected repo type |
| `huggingface hub-repo-details` | `{repo_ids: list[str], repo_type?: "model" \| "dataset" \| "space", operations?: list["overview" \| "dataset_structure" \| "dataset_preview"], config?: str, split?: str, offset?: int, limit?: int}` | Get details for one or more Hugging Face repos (model, dataset, or space). Auto-detects type unless specified. For datasets, use operations: overview, dataset_structure, dataset_preview. Use dataset_structure first to discover configs, splits, sizes, and schema. Use dataset_preview only when config and split are known, unless the dataset has a single config/split. — `repo_ids`: Repo IDs for (models\|dataset/space) - usually in author/name format (e.g. openai/gpt-oss-120b); `repo_type`: Specify lookup type; otherwise auto-detects; `operations`: Details to return. Defaults to ["overview"]. For datasets, prefer ["overview", "dataset_structure"] first; then call ["dataset_preview"] with config and split.; `config`: Dataset Viewer config. Required for dataset_preview when the dataset has multiple config/split options. Discover via dataset_structure.; `split`: Dataset Viewer split. Required for dataset_preview when the dataset has multiple config/split options. Discover via dataset_structure.; `offset`: Row offset for dataset_preview. Defaults to 0.; `limit`: Row count for dataset_preview. Defaults to 5 and is clamped to 1-100. |
| `huggingface hf-fs` | `{operations: list[{cmd: "ls" \| "cat" \| "attach" \| "stat" \| "find" \| "search", args: list[str]}]}` | When to use: Hugging Face Hub models, datasets, Spaces, collections, papers, daily papers, today's trending models, current paper leaderboard, docs, and repository files. Examples: {"operations":[{"cmd":"ls","args":["hf://models/trending","--limit","10"]}]} {"operations":[{"cmd":"ls","args":["hf://papers/trending"]}]} {"operations":[{"cmd":"ls","args":["hf://papers/daily/latest"]}]} Use hf_fs for Hugging Face Hub filesystem operations. Call it with operations, an array of {cmd, args} items; multiple operations may be submitted together. Usage: {"operations":[{"cmd":"ls","args":["hf://models/org/repo"]}]} Grammar; each string below is one args array item: ls URI [--recursive] [--glob GLOB] [--type TYPE] [--sort SORT] [--limit N] cat URI [--offset N] [--max-bytes N] attach URI [--max-bytes N] stat URI find URI [--name GLOB] [--path GLOB] [--type TYPE] [--limit N] search URI [QUERY] [--type TYPE] [--sort SORT] [--tag TAG] [--kind mcp] [--limit N] COMMAND = ls\|cat\|attach\|stat\|find\|search. TYPE = file\|dir\|repo\|bucket\|collection\|paper\|link. SORT = createdAt\|downloads\|likes\|lastModified\|likes30d\|trendingScore\|mainSize\|id\|trending\|upvotes. URI is a canonical hf:// URI. QUERY and GLOB are each one string. Use search for discovery, ls for a known directory, find for recursive matching within a known scope, stat for filesystem metadata or an uncertain target type, cat for text contents, and attach for a complete JPEG, PNG, or WebP image. When the request gives an exact text-file URI, use cat directly; do not add ls or stat first. stat does not read the contents of JSON, Markdown, or other text files. Search scopes: hf://models\|datasets\|spaces[/OWNER], hf://collections[/OWNER], hf://papers, and hf://docs[/...]. Paper and documentation search require QUERY. Repeat --tag only for search hf://spaces; --kind mcp selects MCP Spaces. Use ls hf://models/trending, hf://datasets/trending, hf://spaces/trending, or hf://papers/trending for trending listings. For a named paper.md or metadata.json, use cat directly. Use ls on a paper only to discover an unnamed related VFS. Omit --limit, --sort, and --type unless the request requires them. Limits and path-specific behavior are documented at hf://README.md. Issue one hf_fs call. — `operations.cmd`: Command to execute.; `operations.args`: Command arguments. First item must be an hf:// URI, not a local path or bare filename. One argument per array item. |

## git

Use `git --help` for Git operations on workspace repositories.

## python3

Execute Python code directly under the agent workspace and return stdout
and stderr. Code can be pasted into a `.py` file and run with `python3`.

Python cannot launch subprocesses or make direct network requests. Run the
documented service commands directly in the shell. Use Python for local
file processing and generating JSON arguments.

name	workspace-commands
description	The shell commands and task tools available in this workspace.

Available commands

Standard Unix commands, among them:

cat  head  tail  ls  find  grep  wc  sort  uniq  cut  tr  sed  awk
diff  tee  xargs  jq  unzip  curl

with pipes, > redirection, $(...) substitution and $'...'.

The task CLIs below are registered with this shell rather than installed as files; do not inspect them with which, file or /usr/bin.

Use hf --help for Hugging Face repository, download and upload commands. Repository paths and revisions identify the same files as upstream. Use huggingface for the upstream MCP search and Dataset Viewer operations.

huggingface

Use huggingface <command> '<JSON arguments>'. Run huggingface <command> --help to get more information.

Command	Arguments	Description
`huggingface hf-whoami`	`{}`	Inspect the current Hugging Face authentication context, including the account, visible organization memberships, and credential access details. Read-only and never returns credential values.
`huggingface hub-repo-search`	`{query?: str, repo_types?: list["model" \	"dataset" \	"space"] = ["model", "dataset"], author?: str, filters?: list[str], sort?: "trendingScore" \	"downloads" \	"likes" \	"createdAt" \	"lastModified", limit?: number = 20}`	Search Hugging Face repositories with a shared query interface. You can target models, datasets, spaces, or aggregate across multiple repo types in one call. Include links to repositories in your response. — `query`: Search term. Leave blank and specify sort + limit to browse trending or recent repositories.; `repo_types`: Repository types to search. Defaults to ["model", "dataset"]. space uses keyword search via /api/spaces.; `author`: Organization or user namespace to filter by (e.g. 'google', 'meta-llama', 'huggingface').; `filters`: Optional hub filter tags. Applied to each selected repo type (e.g. ["text-generation"], ["language:en"], ["mcp-server"]).; `sort`: Sort order (descending): trendingScore, downloads, likes, createdAt, lastModified; `limit`: Maximum number of results to return per selected repo type
`huggingface hub-repo-details`	`{repo_ids: list[str], repo_type?: "model" \	"dataset" \	"space", operations?: list["overview" \	"dataset_structure" \	"dataset_preview"], config?: str, split?: str, offset?: int, limit?: int}`	Get details for one or more Hugging Face repos (model, dataset, or space). Auto-detects type unless specified. For datasets, use operations: overview, dataset_structure, dataset_preview. Use dataset_structure first to discover configs, splits, sizes, and schema. Use dataset_preview only when config and split are known, unless the dataset has a single config/split. — `repo_ids`: Repo IDs for (models\	dataset/space) - usually in author/name format (e.g. openai/gpt-oss-120b); `repo_type`: Specify lookup type; otherwise auto-detects; `operations`: Details to return. Defaults to ["overview"]. For datasets, prefer ["overview", "dataset_structure"] first; then call ["dataset_preview"] with config and split.; `config`: Dataset Viewer config. Required for dataset_preview when the dataset has multiple config/split options. Discover via dataset_structure.; `split`: Dataset Viewer split. Required for dataset_preview when the dataset has multiple config/split options. Discover via dataset_structure.; `offset`: Row offset for dataset_preview. Defaults to 0.; `limit`: Row count for dataset_preview. Defaults to 5 and is clamped to 1-100.
`huggingface hf-fs`	`{operations: list[{cmd: "ls" \	"cat" \	"attach" \	"stat" \	"find" \	"search", args: list[str]}]}`	When to use: Hugging Face Hub models, datasets, Spaces, collections, papers, daily papers, today's trending models, current paper leaderboard, docs, and repository files. Examples: {"operations":[{"cmd":"ls","args":["hf://models/trending","--limit","10"]}]} {"operations":[{"cmd":"ls","args":["hf://papers/trending"]}]} {"operations":[{"cmd":"ls","args":["hf://papers/daily/latest"]}]} Use hf_fs for Hugging Face Hub filesystem operations. Call it with operations, an array of {cmd, args} items; multiple operations may be submitted together. Usage: {"operations":[{"cmd":"ls","args":["hf://models/org/repo"]}]} Grammar; each string below is one args array item: ls URI [--recursive] [--glob GLOB] [--type TYPE] [--sort SORT] [--limit N] cat URI [--offset N] [--max-bytes N] attach URI [--max-bytes N] stat URI find URI [--name GLOB] [--path GLOB] [--type TYPE] [--limit N] search URI [QUERY] [--type TYPE] [--sort SORT] [--tag TAG] [--kind mcp] [--limit N] COMMAND = ls\	cat\	attach\	stat\	find\	search. TYPE = file\	dir\	repo\	bucket\	collection\	paper\	link. SORT = createdAt\	downloads\	likes\	lastModified\	likes30d\	trendingScore\	mainSize\	id\	trending\	upvotes. URI is a canonical hf:// URI. QUERY and GLOB are each one string. Use search for discovery, ls for a known directory, find for recursive matching within a known scope, stat for filesystem metadata or an uncertain target type, cat for text contents, and attach for a complete JPEG, PNG, or WebP image. When the request gives an exact text-file URI, use cat directly; do not add ls or stat first. stat does not read the contents of JSON, Markdown, or other text files. Search scopes: hf://models\	datasets\	spaces[/OWNER], hf://collections[/OWNER], hf://papers, and hf://docs[/...]. Paper and documentation search require QUERY. Repeat --tag only for search hf://spaces; --kind mcp selects MCP Spaces. Use ls hf://models/trending, hf://datasets/trending, hf://spaces/trending, or hf://papers/trending for trending listings. For a named paper.md or metadata.json, use cat directly. Use ls on a paper only to discover an unnamed related VFS. Omit --limit, --sort, and --type unless the request requires them. Limits and path-specific behavior are documented at hf://README.md. Issue one hf_fs call. — `operations.cmd`: Command to execute.; `operations.args`: Command arguments. First item must be an hf:// URI, not a local path or bare filename. One argument per array item.

git

Use git --help for Git operations on workspace repositories.

python3

Execute Python code directly under the agent workspace and return stdout and stderr. Code can be pasted into a .py file and run with python3.

Python cannot launch subprocesses or make direct network requests. Run the documented service commands directly in the shell. Use Python for local file processing and generating JSON arguments.

Model	VFS	MCP	VFS Tokens	MCP Tokens	Token Reduction Rate
Claude Opus 5.5	1/1	1/1	238k	267k	11%
GPT-6.1 Sol	1/1	1/1	137k	236k	42%
GPT-5.6 Luna	1/1	1/1	1.47M	5.75M	74%
Claude Haiku 5.5	0/1	1/1	576k	5.85M	90%
Gemini 3.8 Flash	0/1	0/1	3.57M	4.44M	20%

huggingface-upload

Task Prompt

VFS.md

SKILL.md