# Co Scientist Multi Task Learning

> Multi-task learning skill. Shared representation learning, task-specific heads, gradient balancing, auxiliary task selection, and multi-task model evaluation. Use when working with shared representation learning, task-specific heads, gradient balancing.

- Skill: `nahisaho/co-scientist-multi-task-learning` (Agent Skill)
- Install (CLI): `npx skillmds@latest add nahisaho/co-scientist-multi-task-learning`
- Raw SKILL.md: https://api.skillmd.com/api/skills/nahisaho/co-scientist-multi-task-learning/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: nahisaho (https://skillmd.com/u/nahisaho)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/nahisaho/co-scientist-multi-task-learning

---


# Multi-task learning

Multi-task learning skill. Shared representation learning, task-specific heads, gradient balancing, auxiliary task selection, and multi-task model evaluation.

## Use This Skill When

- Shared representation learning.
- Task-specific heads.
- Gradient balancing.
- Auxiliary task selection.
- Multi-task model evaluation.

## Required Inputs

- Research objective, decision target, or hypothesis.
- Available data, source constraints, and domain assumptions.
- Required outputs, success metrics, and deadline or reproducibility constraints.

## Workflow

1. Confirm scope, assumptions, and the exact artifact set to save.
2. Apply the narrowest domain method that answers the request with defensible evidence.
3. Save code, tables, figures, and intermediate outputs to files instead of chat-only output.
4. State limitations, uncertainty, and any validation or sensitivity checks performed.
5. Append skill selection, handoff I/O, and file writes to `logs/process-log.jsonl`.

## Deliverables

- `report.md`: concise method, results, interpretation, and file inventory in the user's language.
- `results/`: structured outputs, metrics, model artifacts, or extracted findings.
- `figures/`: English-only charts, diagrams, or panels when visual output is needed.
- `data/`: processed or derived datasets when transformation occurs.

## Quality Gates

- [ ] The selected method matches the scientific question and stated assumptions.
- [ ] Outputs are reproducible, saved to files, and traceable from inputs to conclusions.
- [ ] Missing data, uncertainty, bias, and hard limits are made explicit.
- [ ] `report.md` and `logs/process-log.jsonl` reference the generated artifacts.
- [ ] No essential result remains chat-only.

If any gate fails: identify the specific failing check, fix the issue, and re-validate before proceeding.

## Gotchas

- Data leakage between train/test splits invalidates all metrics. Verify no leakage before reporting results
- Random seeds must be set for numpy, random, and framework-specific RNGs separately (torch, tf)
- Hyperparameter tuning needs held-out test data never seen during tuning. Three-way split is minimum

## Validation Loop

1. Execute analysis and generate outputs
2. Check:
   - Method selection matches the research question and stated assumptions
   - All outputs are saved to files (no chat-only results)
   - Limitations and uncertainty are explicitly stated
   - `logs/process-log.jsonl` is updated with execution trace
3. If any check fails:
   - Identify the failing gate
   - Fix the specific issue
   - Re-run validation
4. Proceed only after all gates pass