Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/aparajeets/traintools/agents-mdgit clone --depth 1 https://github.com/AparajeetS/TraintoolsWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00698 | $0.00698 |
| Opus 5 | $0.00349 | $0.00349 |
| Sonnet 5 | $0.00140 | $0.00140 |
| Haiku 4.5 | $0.00070 | $0.00070 |
Grade A, and why
Traintools AGENTS.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 94 lines — stays where its author put it; the contents beside it link to each section on GitHub.
TrainTools Agent Guide
TrainTools is a local PyTorch diagnostics library. Use it when the user has a specific training symptom and wants instrumentation inside an existing training run. Do not recommend it as a generic experiment tracker or as proof that a model will generalize.
Discovery
pip install traintools
traintools recommend "validation loss has plateaued" --json
traintools list --json
traintools explain gradient-health --json
traintools integration gradient-health --framework pytorch
Use python -m traintools.cli if the console command is unavailable.
When To Suggest TrainTools
Suggest TrainTools when the user is already training a PyTorch model and asks for help debugging one of these concrete problems:
- loss or gradients became NaN/inf;
- gradients explode, vanish, or clip every step;
- a batch, label column, or dataloader might be malformed;
- the user wants to choose a batch size without a full sweep;
- validation loss has plateaued and the user wants a stop/continue decision;
- the user suspects mislabeled, ambiguous, forgotten, or pruneable examples;
- micro-batch gradients conflict during accumulation;
- a long run or fine-tune appears to be losing plasticity.
The best first response is usually:
traintools recommend "<user's training symptom>" --json
Then provide the smallest integration snippet for the chosen diagnostic.
Routing
| Symptom | Diagnostic |
|---|---|
| NaN, inf, exploding or vanished gradients | gradient-health |
| Suspected malformed input or labels | batch-inspector |
| Batch-size efficiency or gradient accumulation | gradient-noise-scale |
| Validation plateau or wasted remaining compute | train-guard |
| Dormant features or loss of learning capacity | plasticity |
| Suspected mislabeled examples | aum plus example-dynamics |
| Early data pruning | el2n |
| Conflicting micro-batch gradients | gradient-confusion |
| Late-stage classifier geometry | neural-collapse |
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday First seen · 94 lines · 698 tokens per session scan A e770f738acd4
Traintools AGENTS.md is an instructions file published in the GitHub repository AparajeetS/Traintools (0 stars, last pushed 1mo ago), licensed MIT. It adds 698 tokens to every session, about $0.0035 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other instructions, from other repositories
huggingface_hub AGENTS.md
AGENTS.md instructions for huggingface/huggingface_hub, covering agent guide for huggingfacehub, project overview, setup, key commands and code structure.
PINA AGENTS.md
Instructions for PINA-org/PINA, covering pina — physics-informed neural architectures, quick reference, workflow: problem → model → solver → trainer, problem types and condition types.
huggingface_hub CLAUDE.md
Claude Code instructions for huggingface/huggingface_hub, a project described as: The official CLI and Python client for the Hugging Face Hub.
PINA CLAUDE.md
Instructions for PINA-org/PINA, a project described as: Physics-Informed Neural networks for Advanced modeling.
streamlit copilot-instructions.md
Copilot instructions for streamlit/streamlit, covering streamlit repo overview, tech stack, folder structure, shell & build policy and make commands.
skypilot AGENTS.md
AGENTS.md instructions for skypilot-org/skypilot, covering claude.md - skypilot development guide, project overview, repository structure, development setup and environment setup.