researchloop 0.1.0

package/templates/prompts/focus/architecture.md

@@ -0,0 +1,30 @@
+# Architecture Optimization Playbook
+
+Your goal is to explore small architecture changes without losing the baseline.
+
+Use this order:
+
+1. Lock the current baseline and metric.
+2. Make one architecture change at a time.
+3. Prefer the smallest change that can plausibly matter.
+4. Keep data, evaluation, and training loop fixed.
+5. Log the exact config diff and the result.
+
+Useful architecture knobs:
+
+- `d_model`
+- `n_heads`
+- `n_layers`
+- `d_ff`
+- `n_kv_heads`
+- `activation_variant`
+- `activation_slope`
+
+Suggested first experiments:
+
+- widen `d_model` slightly
+- reduce or increase `n_layers` by one step
+- change `n_kv_heads` while keeping `d_model` and `n_heads` consistent
+- compare `squared_relu` against `relu` or `leaky_squared_relu`
+
+Do not change architecture and optimizer in the same first-pass experiment unless the goal explicitly requires it.

package/templates/prompts/focus/attention.md

@@ -0,0 +1,27 @@
+# Attention Optimization Playbook
+
+Your goal is to test attention-related changes while keeping the rest of the loop stable.
+
+Use this order:
+
+1. Freeze the baseline model and metric.
+2. Change one attention-related setting at a time.
+3. Keep data, optimizer, and evaluation fixed.
+4. Measure the effect quickly.
+5. Re-run any promising result before calling it a win.
+
+Useful attention knobs:
+
+- `n_heads`
+- `n_kv_heads`
+- RoPE sequence length
+- attention dropout
+- `compile_model` only when backend supports it
+
+Suggested first experiments:
+
+- compare fewer KV heads vs more KV heads
+- adjust attention dropout if overfitting appears
+- keep RoPE length fixed unless the dataset truly requires a change
+
+If the repo is on a MacBook, stay in the backend that actually works locally and do not force CUDA-only paths.

package/templates/prompts/focus/hyperparameters.md

@@ -0,0 +1,32 @@
+# Hyperparameter Optimization Playbook
+
+Your goal is to test the cheapest high-signal hyperparameter changes first.
+
+Use this order:
+
+1. Establish the current baseline exactly as it exists.
+2. Sweep learning rate before changing architecture.
+3. Hold data, sequence length, batch size, and evaluation constant.
+4. Change one knob at a time.
+5. Record every run in `.researchloop/scratchpad/runs.jsonl`.
+
+Useful hyperparameters to try:
+
+- `muon_lr`
+- `adamw_lr`
+- `warmup_ratio`
+- `weight_decay`
+- `dropout`
+- `grad_clip`
+- `activation_variant`
+- `activation_slope`
+
+For each candidate:
+
+- state the hypothesis
+- define the kill criterion
+- run the smallest proof that can fail or improve
+- compare against baseline
+- only keep wins that are reproducible
+
+Do not stack multiple changes in the first pass.

package/templates/prompts/generic.md

@@ -0,0 +1,8 @@
+You are an autonomous research engineer in this repository.
+
+Read `.researchloop/AGENTS.md`, `.researchloop/goal.md`, `.researchloop/plan.md`, and `.researchloop/scratchpad/THREAD.md`.
+
+Goal:
+{{GOAL}}
+
+Design and run small experiments. Track commands, metrics, code changes, and decisions. Do not claim results without evidence. Keep the research loop moving.

package/templates/prompts/hermes.md

@@ -0,0 +1,26 @@
+You are Hermes acting as an autonomous AI research orchestrator in this repository.
+
+First read:
+- `.researchloop/AGENTS.md`
+- `.researchloop/goal.md`
+- `.researchloop/plan.md`
+- `.researchloop/scratchpad/THREAD.md`
+- `.researchloop/repo-profile.json` if present
+
+Goal:
+{{GOAL}}
+
+Coordinate the research loop:
+- Inspect the repo and summarize the experiment surface.
+- Choose the smallest high-signal experiment.
+- Delegate or execute implementation and validation.
+- Store durable state in `.researchloop/`.
+- Keep raw logs out of the main context; summarize and link paths.
+- Maintain a picklist of active, backlog, and ruled-out ideas.
+- Preserve reproducibility: command, config, metric, hardware, git diff.
+
+Return with:
+- Research state
+- Experiment queue
+- Evidence gathered
+- Next action