polygramme-cloud

The Polygramme Cloud client. Train LoRA adapters on a hosted trainer, chat with them the moment a checkpoint lands, and download the weights — from any laptop. No torch, no GPU dependencies.

pip install polygramme-cloud
export POLYGRAMME_API_KEY=pg_…          # https://polygramme.com/console → API keys

polygramme train --dataset conversations.jsonl --save-every 25
polygramme chat --model <model_id> "hi"
polygramme checkpoints
polygramme download checkpoint/<model_id>/final.tar.gz
polygramme export <model_id>              # merged full HF checkpoint

conversations.jsonl is one {"messages": [{"role": ..., "content": ...}, ...]} per line (what rlcli import produces). The shared trainer is Qwen3.5-9B with rank-32 LoRA adapters.

Python:

from polygramme_cloud import Cloud
c = Cloud()                                   # POLYGRAMME_API_KEY from env
for delta in c.chat("<model_id>", [{"role": "user", "content": "hi"}], stream=True):
    print(delta, end="")

For training on your own GPUs, use rlcli, the open framework this cloud runs.

Docs and the agent skill

The Guide, Reference and an Agent Skills file for this stack are customer docs: served only with your API key, never bundled here.

polygramme skill --install          # writes .claude/skills/polygramme-cloud/SKILL.md (--global for ~/.claude/skills)
polygramme docs guide               # or reference / skill — prints Markdown

The install verifies the file against the checksum in GET /api/docs/manifest and never writes your key anywhere.