Advanced track
Fine-tune a model too large for one GPU. You score Kev-27B on one Yens H200 to set the bar, fine-tune Qwen3.8-27B on Marlowe on one node and then two, serve Qwen 235B across a full node, and distill a small student from a teacher’s labels.
| Task | Where | What you do |
|---|---|---|
| A1. Score Kev-27B before training | Yens, 1 H200 | Serve Kev-27B and score it on 1,000 test filings |
| A2. Fine-tune Qwen 27B | Marlowe, 1 H100, then 8 H100s | Fine-tune it on one GPU, then on 8 with FSDP, and compare |
| A3. Fine-tune on two nodes | Marlowe, 16 H100s | Run the same fine-tune across 2 nodes and compare with 1 node |
| A4. Inference at scale with Qwen 235B | Marlowe, 8 H100s | Load Qwen 235B on one node, time the weight load, and label filings |
| A5. Distill a student | Marlowe, 1 H100 | Train Qwen 4B on teacher labels (yours or DARC’s) and score it |
Data and models are in place on both clusters before the day, so no task waits on a transfer. If a step runs late, DARC’s result for it lets you go on.
Data and models
| Task | Data | Model |
|---|---|---|
| A1 | 1,000 test filings (2022 to 2026) with Jev’s labels | Kev-27B as released |
| A2 | train.jsonl: 54,426 filings (2004 to 2018) with Jev’s soft labels |
Qwen3.8-27B |
| A3 | train.jsonl |
Qwen3.8-27B |
| A4 | the training filings, no labels needed | Qwen 235B |
| A5 | the training filings with teacher labels; the 1,000 test filings to score | Qwen3.5-4B-Base |
Bonus
- Move a run between the Yens and Marlowe with Globus.
- Run the distilled student on the Yens.
The Advanced track has not had its dry run yet. A2 to A5 are still being written.