EvoScientist 0.0.1.dev2__py3-none-any.whl

EvoScientist/middleware.py

@@ -0,0 +1,35 @@
+"""Middleware configuration for the EvoScientist agent."""
+
+from pathlib import Path
+
+from deepagents.middleware.skills import SkillsMiddleware
+
+from .backends import MergedReadOnlyBackend
+
+_DEFAULT_SKILLS_DIR = str(Path(__file__).parent / "skills")
+
+
+def create_skills_middleware(
+    skills_dir: str = _DEFAULT_SKILLS_DIR,
+    workspace_dir: str = "./workspace/",
+) -> SkillsMiddleware:
+    """Create a SkillsMiddleware that loads skills.
+
+    Merges user-installed skills (workspace/skills/) with system skills
+    (package built-in). User skills take priority on name conflicts.
+
+    Args:
+        skills_dir: Path to the system skills directory (package built-in)
+        workspace_dir: Path to the workspace root (user skills live under workspace/skills/)
+
+    Returns:
+        Configured SkillsMiddleware instance
+    """
+    merged = MergedReadOnlyBackend(
+        primary_dir=str(Path(workspace_dir) / "skills"),
+        secondary_dir=skills_dir,
+    )
+    return SkillsMiddleware(
+        backend=merged,
+        sources=["/"],
+    )

EvoScientist/prompts.py

@@ -0,0 +1,277 @@
+"""Prompt templates for the EvoScientist experimental agent."""
+
+# =============================================================================
+# Main agent workflow
+# =============================================================================
+
+EXPERIMENT_WORKFLOW = """# Experiment Workflow
+
+You are the main experimental agent. Your mission is to transform a research proposal
+into reproducible experiments and a paper-ready experimental report.
+
+## Core Principles
+- Baseline first, then iterate (ablation-friendly).
+- Change one major variable per iteration (data, model, objective, or training recipe).
+- Never invent results. If you cannot run something, say so and propose the smallest next step.
+- Delegate aggressively using the `task` tool. Prefer the research sub-agent for web search.
+- Use local skills via `load_skill` when they match the task. Skills provide proven workflows and checklists.
+  All skills are available under `/skills/` (read-only).
+  When calling `load_skill`, use the skill id from the SKILL.md frontmatter (`name:`), not the folder name.
+
+## Scientific Rigor Checklist
+- Validate data and run quick EDA; document anomalies or data leakage risks.
+- Separate exploratory vs confirmatory analyses; define primary metrics up front.
+- Report effect sizes with uncertainty (confidence intervals/error bars) where possible.
+- Apply multiple-testing correction when comparing many conditions.
+- State limitations, negative results, and sensitivity to key parameters.
+- Track reproducibility (seeds, versions, configs, and exact commands).
+
+## Step 1: Intake & Scope
+- Read the proposal and extract goals, datasets, constraints, and evaluation metrics
+- Capture key assumptions and open questions
+- Save the original proposal to `/research_request.md`
+
+## Step 2: Plan (Recommended Structure)
+- Create experiment stages with success signals (flexible, not rigid)
+- Identify resource/data dependencies and baseline requirements
+- Use `write_todos` to track the execution plan and updates
+- If delegating planning to planner-agent, start your message with: `MODE: PLAN`
+- If a stage matches an existing skill, note the skill name in the plan and load it before implementation.
+  Use the skill id from SKILL.md frontmatter (`name:`).
+-- Save the plan to `/todos.md` (recommended). Include per-stage:
+  - objective and success signals
+  - what to run (commands/scripts)
+  - expected artifacts (tables/plots/logs)
+- Optionally save:
+  - `/plan.md` for stages
+  - `/success_criteria.md` for success signals
+
+## Step 3: Execute & Debug
+- Delegate tasks to sub-agents using the `task` tool:
+  - Planning/structuring → planner-agent
+  - Methods/baselines/datasets → research-agent
+  - Implementation → code-agent
+  - Debugging → debug-agent
+  - Analysis/visualization → data-analysis-agent
+  - Report drafting → writing-agent
+- Prefer the research-agent for web search; avoid searching directly
+- Use `execute` for shell commands when running experiments
+- When a task matches an existing skill, `load_skill` it and follow it rather than reinventing the workflow.
+- Keep outputs organized under `/artifacts/` (recommended)
+- Optionally log runs to `/experiment_log.md` (params, seeds, env, outputs)
+
+## Step 4: Evaluate & Iterate
+- Compare results against success signals
+- If results are weak or ambiguous, iterate:
+  - identify gaps
+  - propose new methods/data
+  - re-run and re-evaluate
+- Prefer evidence-driven iteration: error analysis, sanity checks, and minimal ablations
+- Update `/todos.md` to reflect new iterations
+- Stop iterating when evidence is sufficient or diminishing returns appear
+
+### Stage Reflection (Recommended Checkpoint)
+After any meaningful experimental stage (baseline, new dataset, new training recipe, etc.),
+delegate a short reflection to the planner-agent and use it to update the remaining plan.
+
+Trigger this checkpoint when:
+- A baseline finishes (you now have a reference point).
+- You introduce a new dataset/model/training recipe (risk of confounding changes).
+- Two iterations in a row fail to improve the primary metric.
+- Results look suspicious (metric mismatch, unstable training, unexpected regressions).
+
+When calling the planner-agent in reflection mode, provide:
+- Start your message with: `MODE: REFLECTION`
+- Stage name/index and intent
+- Commands run + key parameters (model, dataset, seeds, batch size, lr, epochs, hardware)
+- Key metrics vs baseline (a small table is ideal)
+- Artifact paths (logs, plots, checkpoints)
+- Which success signals were met/unmet
+- If proposing skills, use skill ids from SKILL.md frontmatter (`name:`).
+
+Ask the planner-agent to output a **Plan Update JSON** with this schema:
+```json
+{
+  "completed": ["..."],
+  "unmet_success_signals": ["..."],
+  "skill_suggestions": ["..."],
+  "stage_modifications": [
+    {"stage": "Stage name or index", "change": "What to adjust and why"}
+  ],
+  "new_stages": [
+    {
+      "title": "...",
+      "goal": "...",
+      "success_signals": ["..."],
+      "what_to_run": ["..."],
+      "expected_artifacts": ["..."]
+    }
+  ],
+  "todo_updates": ["..."]
+}
+```
+Empty arrays are valid. If no changes are needed, return the JSON with empty arrays.
+Then revise `/todos.md` accordingly.
+
+## Step 5: Write Report
+- Write the final report to `/final_report.md` (Markdown)
+- Include:
+  - Problem summary
+  - Experiment plan (stages + success signals)
+  - Experimental setup and configurations
+  - Results and visualizations (reference artifacts)
+  - Analysis, limitations, and next steps
+- If web research was used, include a Sources section with real URLs (no fabricated citations)
+- When applicable, include effect sizes, uncertainty, and notes on statistical corrections.
+- Be precise, technical, and concise
+
+## Step 6: Verify
+- Re-read `/research_request.md` to ensure coverage
+- Confirm the report answers the proposal and documents key settings/results
+
+## Experiment Report Template (Recommended)
+1. Summary & goals
+2. Experiment plan (stages + success signals)
+3. Setup (data, model, environment, parameters)
+4. Baselines and comparisons
+5. Results (tables/figures + references to artifacts)
+6. Analysis, limitations, and next steps
+
+## Writing Guidelines
+- Use bullets for configs, stage lists, and key results; use short paragraphs for reasoning
+- Avoid first-person singular ("I ..."). Prefer neutral phrasing ("This experiment...") or "we" style.
+- Professional, objective tone
+
+## Shell Execution Guidelines
+When using the `execute` tool for shell commands:
+
+**Short commands** (< 30 seconds): Run directly
+```bash
+python script.py
+pip install pandas
+```
+
+**Long-running commands** (> 30 seconds): Run in background, then check results
+```bash
+# Step 1: Start in background, redirect output to log
+python long_task.py > /output.log 2>&1 &
+
+# Step 2: Check if still running
+ps aux | grep long_task
+
+# Step 3: Read results when done
+cat /output.log
+```
+
+This prevents blocking the conversation during long operations.
+"""
+
+# =============================================================================
+# Sub-agent delegation strategy
+# =============================================================================
+
+DELEGATION_STRATEGY = """# Sub-Agent Delegation
+
+## Default: Use 1 Sub-Agent
+For most tasks, a single sub-agent is sufficient:
+- "Plan experimental stages" → planner-agent
+- "Reflect and update the plan after a stage" → planner-agent
+- "Find related methods/baselines/datasets" → research-agent
+- "Implement baseline or training loop" → code-agent
+- "Debug runtime failures" → debug-agent
+- "Analyze metrics and plot figures" → data-analysis-agent
+- "Draft report sections" → writing-agent
+
+## Task Granularity
+- One sub-agent task = one topic / one experiment / one artifact bundle
+- Provide concrete file paths, commands, and success signals in each task
+  so the sub-agent can respond precisely
+
+## Parallelize Only When Necessary
+Use multiple sub-agents ONLY for:
+
+**Explicit comparisons** (1 per method/baseline):
+- "Compare A vs B vs C" → 3 parallel sub-agents
+
+**Distinct experiments** with separate datasets or setups:
+- "Run baselines on X and Y" → 2 parallel sub-agents
+
+## Limits
+- Maximum {max_concurrent} parallel sub-agents per round
+- Maximum {max_iterations} delegation rounds total
+- Stop when evidence is sufficient
+
+## Key Principles
+- Bias towards a single sub-agent (token-efficient)
+- Avoid premature decomposition
+- Each sub-agent returns focused, self-contained findings
+"""
+
+# =============================================================================
+# Sub-agent research instructions
+# =============================================================================
+
+RESEARCHER_INSTRUCTIONS = """You are a research assistant. Today's date is {date}.
+
+## Task
+Use tools to gather information on the assigned topic (methods, baselines,
+datasets, or prior results) to support experimental planning or iteration.
+Prefer actionable details: datasets, metrics, code availability, and common pitfalls.
+Do not fabricate citations or URLs.
+Capture evaluation protocols (splits, metrics, calibration) and known failure modes.
+
+## Available Tools
+1. **tavily_search** - Web search for information
+2. **think_tool** - Reflect on findings and plan next steps
+
+**CRITICAL: Use think_tool after each search**
+
+## Research Strategy
+1. Read the question carefully
+2. Start with broad searches
+3. After each search, reflect: Do I have enough? What's missing?
+4. Narrow searches to fill gaps
+5. Stop when you can answer confidently
+
+## Hard Limits
+- Simple queries: 2-3 searches maximum
+- Complex queries: up to 5 searches maximum
+- Stop after 5 searches regardless
+
+## Stop When
+- You can answer comprehensively
+- You have 3+ relevant sources
+- Last 2 searches returned similar information
+
+## Response Format
+Structure findings with clear headings and cite sources inline:
+
+```
+## Key Findings
+
+Finding one with context [1]. Another insight [2].
+
+## Recommended Next Experiments
+- One actionable experiment suggestion with motivation and expected outcome.
+
+### Sources
+[1] Title: URL
+[2] Title: URL
+```
+"""
+
+# =============================================================================
+# Combined exports
+# =============================================================================
+
+def get_system_prompt(max_concurrent: int = 3, max_iterations: int = 3) -> str:
+    """Generate the complete system prompt with configured limits."""
+    delegation = DELEGATION_STRATEGY.format(
+        max_concurrent=max_concurrent,
+        max_iterations=max_iterations,
+    )
+    return EXPERIMENT_WORKFLOW + "\n" + delegation
+
+
+# Default export (backward compatible)
+SYSTEM_PROMPT = get_system_prompt()

EvoScientist/skills/accelerate/SKILL.md

@@ -0,0 +1,332 @@
+---
+name: accelerate
+description: Simplest distributed training API. 4 lines to add distributed support to any PyTorch script. Unified API for DeepSpeed/FSDP/Megatron/DDP. Automatic device placement, mixed precision (FP16/BF16/FP8). Interactive config, single launch command. HuggingFace ecosystem standard.
+version: 1.0.0
+author: Orchestra Research
+license: MIT
+tags: [Distributed Training, HuggingFace, Accelerate, DeepSpeed, FSDP, Mixed Precision, PyTorch, DDP, Unified API, Simple]
+dependencies: [accelerate, torch, transformers]
+---
+
+# HuggingFace Accelerate - Unified Distributed Training
+
+## Quick start
+
+Accelerate simplifies distributed training to 4 lines of code.
+
+**Installation**:
+```bash
+pip install accelerate
+```
+
+**Convert PyTorch script** (4 lines):
+```python
+import torch
++ from accelerate import Accelerator
+
++ accelerator = Accelerator()
+
+  model = torch.nn.Transformer()
+  optimizer = torch.optim.Adam(model.parameters())
+  dataloader = torch.utils.data.DataLoader(dataset)
+
++ model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
+
+  for batch in dataloader:
+      optimizer.zero_grad()
+      loss = model(batch)
+-     loss.backward()
++     accelerator.backward(loss)
+      optimizer.step()
+```
+
+**Run** (single command):
+```bash
+accelerate launch train.py
+```
+
+## Common workflows
+
+### Workflow 1: From single GPU to multi-GPU
+
+**Original script**:
+```python
+# train.py
+import torch
+
+model = torch.nn.Linear(10, 2).to('cuda')
+optimizer = torch.optim.Adam(model.parameters())
+dataloader = torch.utils.data.DataLoader(dataset, batch_size=32)
+
+for epoch in range(10):
+    for batch in dataloader:
+        batch = batch.to('cuda')
+        optimizer.zero_grad()
+        loss = model(batch).mean()
+        loss.backward()
+        optimizer.step()
+```
+
+**With Accelerate** (4 lines added):
+```python
+# train.py
+import torch
+from accelerate import Accelerator  # +1
+
+accelerator = Accelerator()  # +2
+
+model = torch.nn.Linear(10, 2)
+optimizer = torch.optim.Adam(model.parameters())
+dataloader = torch.utils.data.DataLoader(dataset, batch_size=32)
+
+model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)  # +3
+
+for epoch in range(10):
+    for batch in dataloader:
+        # No .to('cuda') needed - automatic!
+        optimizer.zero_grad()
+        loss = model(batch).mean()
+        accelerator.backward(loss)  # +4
+        optimizer.step()
+```
+
+**Configure** (interactive):
+```bash
+accelerate config
+```
+
+**Questions**:
+- Which machine? (single/multi GPU/TPU/CPU)
+- How many machines? (1)
+- Mixed precision? (no/fp16/bf16/fp8)
+- DeepSpeed? (no/yes)
+
+**Launch** (works on any setup):
+```bash
+# Single GPU
+accelerate launch train.py
+
+# Multi-GPU (8 GPUs)
+accelerate launch --multi_gpu --num_processes 8 train.py
+
+# Multi-node
+accelerate launch --multi_gpu --num_processes 16 \
+  --num_machines 2 --machine_rank 0 \
+  --main_process_ip $MASTER_ADDR \
+  train.py
+```
+
+### Workflow 2: Mixed precision training
+
+**Enable FP16/BF16**:
+```python
+from accelerate import Accelerator
+
+# FP16 (with gradient scaling)
+accelerator = Accelerator(mixed_precision='fp16')
+
+# BF16 (no scaling, more stable)
+accelerator = Accelerator(mixed_precision='bf16')
+
+# FP8 (H100+)
+accelerator = Accelerator(mixed_precision='fp8')
+
+model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
+
+# Everything else is automatic!
+for batch in dataloader:
+    with accelerator.autocast():  # Optional, done automatically
+        loss = model(batch)
+    accelerator.backward(loss)
+```
+
+### Workflow 3: DeepSpeed ZeRO integration
+
+**Enable DeepSpeed ZeRO-2**:
+```python
+from accelerate import Accelerator
+
+accelerator = Accelerator(
+    mixed_precision='bf16',
+    deepspeed_plugin={
+        "zero_stage": 2,  # ZeRO-2
+        "offload_optimizer": False,
+        "gradient_accumulation_steps": 4
+    }
+)
+
+# Same code as before!
+model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
+```
+
+**Or via config**:
+```bash
+accelerate config
+# Select: DeepSpeed → ZeRO-2
+```
+
+**deepspeed_config.json**:
+```json
+{
+    "fp16": {"enabled": false},
+    "bf16": {"enabled": true},
+    "zero_optimization": {
+        "stage": 2,
+        "offload_optimizer": {"device": "cpu"},
+        "allgather_bucket_size": 5e8,
+        "reduce_bucket_size": 5e8
+    }
+}
+```
+
+**Launch**:
+```bash
+accelerate launch --config_file deepspeed_config.json train.py
+```
+
+### Workflow 4: FSDP (Fully Sharded Data Parallel)
+
+**Enable FSDP**:
+```python
+from accelerate import Accelerator, FullyShardedDataParallelPlugin
+
+fsdp_plugin = FullyShardedDataParallelPlugin(
+    sharding_strategy="FULL_SHARD",  # ZeRO-3 equivalent
+    auto_wrap_policy="TRANSFORMER_AUTO_WRAP",
+    cpu_offload=False
+)
+
+accelerator = Accelerator(
+    mixed_precision='bf16',
+    fsdp_plugin=fsdp_plugin
+)
+
+model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
+```
+
+**Or via config**:
+```bash
+accelerate config
+# Select: FSDP → Full Shard → No CPU Offload
+```
+
+### Workflow 5: Gradient accumulation
+
+**Accumulate gradients**:
+```python
+from accelerate import Accelerator
+
+accelerator = Accelerator(gradient_accumulation_steps=4)
+
+model, optimizer, dataloader = accelerator.prepare(model, optimizer, dataloader)
+
+for batch in dataloader:
+    with accelerator.accumulate(model):  # Handles accumulation
+        optimizer.zero_grad()
+        loss = model(batch)
+        accelerator.backward(loss)
+        optimizer.step()
+```
+
+**Effective batch size**: `batch_size * num_gpus * gradient_accumulation_steps`
+
+## When to use vs alternatives
+
+**Use Accelerate when**:
+- Want simplest distributed training
+- Need single script for any hardware
+- Use HuggingFace ecosystem
+- Want flexibility (DDP/DeepSpeed/FSDP/Megatron)
+- Need quick prototyping
+
+**Key advantages**:
+- **4 lines**: Minimal code changes
+- **Unified API**: Same code for DDP, DeepSpeed, FSDP, Megatron
+- **Automatic**: Device placement, mixed precision, sharding
+- **Interactive config**: No manual launcher setup
+- **Single launch**: Works everywhere
+
+**Use alternatives instead**:
+- **PyTorch Lightning**: Need callbacks, high-level abstractions
+- **Ray Train**: Multi-node orchestration, hyperparameter tuning
+- **DeepSpeed**: Direct API control, advanced features
+- **Raw DDP**: Maximum control, minimal abstraction
+
+## Common issues
+
+**Issue: Wrong device placement**
+
+Don't manually move to device:
+```python
+# WRONG
+batch = batch.to('cuda')
+
+# CORRECT
+# Accelerate handles it automatically after prepare()
+```
+
+**Issue: Gradient accumulation not working**
+
+Use context manager:
+```python
+# CORRECT
+with accelerator.accumulate(model):
+    optimizer.zero_grad()
+    accelerator.backward(loss)
+    optimizer.step()
+```
+
+**Issue: Checkpointing in distributed**
+
+Use accelerator methods:
+```python
+# Save only on main process
+if accelerator.is_main_process:
+    accelerator.save_state('checkpoint/')
+
+# Load on all processes
+accelerator.load_state('checkpoint/')
+```
+
+**Issue: Different results with FSDP**
+
+Ensure same random seed:
+```python
+from accelerate.utils import set_seed
+set_seed(42)
+```
+
+## Advanced topics
+
+**Megatron integration**: See [references/megatron-integration.md](references/megatron-integration.md) for tensor parallelism, pipeline parallelism, and sequence parallelism setup.
+
+**Custom plugins**: See [references/custom-plugins.md](references/custom-plugins.md) for creating custom distributed plugins and advanced configuration.
+
+**Performance tuning**: See [references/performance.md](references/performance.md) for profiling, memory optimization, and best practices.
+
+## Hardware requirements
+
+- **CPU**: Works (slow)
+- **Single GPU**: Works
+- **Multi-GPU**: DDP (default), DeepSpeed, or FSDP
+- **Multi-node**: DDP, DeepSpeed, FSDP, Megatron
+- **TPU**: Supported
+- **Apple MPS**: Supported
+
+**Launcher requirements**:
+- **DDP**: `torch.distributed.run` (built-in)
+- **DeepSpeed**: `deepspeed` (pip install deepspeed)
+- **FSDP**: PyTorch 1.12+ (built-in)
+- **Megatron**: Custom setup
+
+## Resources
+
+- Docs: https://huggingface.co/docs/accelerate
+- GitHub: https://github.com/huggingface/accelerate
+- Version: 1.11.0+
+- Tutorial: "Accelerate your scripts"
+- Examples: https://github.com/huggingface/accelerate/tree/main/examples
+- Used by: HuggingFace Transformers, TRL, PEFT, all HF libraries
+
+
+