Who it is for
Unified Efficient Fine-Tuning of 100+ LLMs
pyproject.toml:8
PUBLIC REPORT · ADDED BY METAL MANTRA / AGENT SIGNAL v0.2
llamafactory.
Scanned repository says: Unified Efficient Fine-Tuning of 100+ LLMs
Metal Mantra scanned this public repository on its own initiative; the maintainer did not submit it. Maintainers: scan the repository yourself to take this report over, or contact us.
Inspect GitHub source0community interest
Votes show community interest, not safety. They never change this score.
77/100BBBNot security certification
Runtime safety, real-world agent capability, fitness for your use case or independent certification.
Scan scope200 / 226Partial eligible-file selection · not the entire repository
Recorded findings7Heuristic matches requiring human review
SnapshotOct 12, 2026UTC · engine 0.4.0
DNS at scan timeUnverifiedDomain control is not security certification
What the source
actually tells us.
Structured, cited information from the scanned archive. Repository-authored claims are labeled and never treated as verified capabilities.
200 / 226Partial selection
ReadRepository-authored information
2LICENSE, pyproject.toml
What it can do
Various models: LLaMA, LLaVA, Mistral, Mixtral-MoE, Qwen3, Qwen3-VL, DeepSeek, Gemma, GLM, Phi, etc.
README.md:105For the "base" models, the template argument can be chosen from default, alpaca, vicuna etc. But make sure to use the corresponding template for the "instruct/chat" models.
README.md:344Integrated methods: (Continuous) pre-training, (multimodal) supervised fine-tuning, reward modeling, PPO, DPO, KTO, ORPO, etc.
README.md:106Scalable resources: 16-bit full-tuning, freeze-tuning, LoRA and 2/3/4/5/6/8-bit QLoRA via AQLM/AWQ/GPTQ/LLM.int8/HQQ/EETQ.
README.md:107
What adoption takes
Installation is mandatory.
README.md:521Manifest declares Python >=3.11.0 as a requirement.
pyproject.toml:11Declares a dependency on torch.
pyproject.toml:40Use the following 3 commands to run LoRA fine-tuning, inference and merging of the Qwen3-4B-Instruct model, respectively.
README.md:677Declares a dependency on torchvision.
pyproject.toml:41Declares a dependency on torchaudio.
pyproject.toml:42Declares a dependency on transformers.
pyproject.toml:43Declares a dependency on datasets.
pyproject.toml:44Declares a dependency on accelerate.
pyproject.toml:45
Limits and terms
A license or copying file is present; its terms have not been reviewed.
LICENSE:1
Observed technical context
Python · 200 selected filesTool-calling reference · static only
These are references in selected source files, not proof that a feature works at runtime.
Read from commit 1fded91 at Oct 12, 2026, 5:13 AM UTC. Source links open that exact commit. Repository statements are unreviewed and do not affect the score.
02 / FINDINGS FIRSTWhat needs
What needs
a closer look.
Critical details are withheld publicly for the first 30 days after a scan. A matched pattern is a lead for investigation, not a confirmed vulnerability.
Dependencies: no lockfile or exact version pins at the repository root, so no advisory check was possible.
7 findings
MAJORUnsafe deserialization
Unsafe deserialization
The code loads data with a format that can run code while it is read, such as pickle or unrestricted YAML. A poisoned file, cache entry or downloaded model then runs with the agent's permissions.
Use a data-only format (JSON, safetensors, yaml.safe_load) for anything that is not fully trusted.
MAJORUnsafe deserialization
Unsafe deserialization
The code loads data with a format that can run code while it is read, such as pickle or unrestricted YAML. A poisoned file, cache entry or downloaded model then runs with the agent's permissions.
Use a data-only format (JSON, safetensors, yaml.safe_load) for anything that is not fully trusted.
MAJORRemote model code trusted
Remote model code trusted
The code loads a model with trust_remote_code enabled, which downloads and runs Python from the model's repository. Whoever controls that repository controls code running inside the agent.
Leave trust_remote_code off, or pin the model to a reviewed revision hash.
MAJORDomain ownership not verified
Domain ownership not verified
Metal Mantra could not confirm that the owner controls the stated domain. The domain may belong to someone else.
Verify the domain using the Metal Mantra DNS TXT challenge.
MAJORHTTPS not confirmed
HTTPS not confirmed
The stated website could not be reached over HTTPS, so Metal Mantra could not confirm it is a secure, working site.
Make the official website reachable over HTTPS for verification.
MINORPrivacy Policy link not confirmed
Privacy Policy link not confirmed
No privacy policy link was found on the stated homepage. Buyers cannot see how the vendor handles their data.
Expose a clear Privacy Policy link on the official homepage.
MINORTerms link not confirmed
Terms link not confirmed
No terms link was found on the stated homepage. Buyers cannot see the conditions that apply to using the service.
Expose a clear Terms link on the official homepage.
The boundary
is part of the result.
- What was scanned
- 200 of 226 eligible files · Partial selection.
Supported static source patterns and domain checks; no runtime execution. - What remains unknown
- Unscanned files, external dependencies, production configuration, actual agent behavior and evidence beyond the configured checks.
- Source location
- Links open the current repository HEAD. Source may have changed since this snapshot.
- Something looks wrong?
- Tell us if a finding, scope or claim on this report needs a correction. Report a problem with this report
- Entity trust · 15% of v0.2
- Includes DNS ownership, HTTPS and policy evidence. Verification is domain control, not a security certificate.DNS: unverifiedHTTPS: not confirmedPrivacy: not confirmedTerms: not confirmed
Open the
calculation.
Each pillar starts at 100. Rule deductions reduce its score; weights determine the overall contribution.
Security & privacy
68/100
Guardrails & loop control
100/100
Schema & tooling
100/100
Token & cost efficiency
100/100
Entity trust
0/100
Did llamafactory work for people?
No user reports yet.
Each report is one signed-in GitHub account saying what happened. We do not verify that the account ran the agent, and reports never change the automated score or grade. For a security problem, use the disclosure page instead of a public note.
SHARE THE EVIDENCEA signal.
A signal.
Not a seal.
Share the full report so its scope and limits travel with the score.
