Framework · experimental

karpathy/autoresearch

Minimal autonomous LLM research loop: agent edits train.py, runs fixed five-minute experiments, checks val_bpb, and keeps/discards changes.

LLM researchautonomous experimentation
Installation GuideInstruction Manual

What it contains

Minimal autonomous LLM research loop: agent edits train.py, runs fixed five-minute experiments, checks val_bpb, and keeps/discards changes.

Typical uses

Let a coding agent explore model/optimizer/training changes on a bounded single-GPU setup while the human controls program.md.

Why people choose it

Inference: chosen for tightly scoped autonomous experiments with a simple metric and reviewable diffs.

License
MIT
Maintenance
active
Latest release
Unknown
Last meaningful update
Unknown
Maturity
experimental
Production readiness
prototype

Classification

Subject domains

LLM researchautonomous experimentation

Task categories

LLM researchautonomous experimentation

Project phases

researchimplementation

Secondary types

experiment harnessagent workflow

Select when

  • Need autonomous small-LLM experiments on suitable NVIDIA GPU

Do not select when

  • General SWE
  • No suitable GPU
  • Need cross-machine metric comparability
Implementation complexity
medium
Setup effort
medium
Learning curve
medium
Confidence
high

Routing heuristic 71.0%

Routing specificity
3/5
Setup simplicity
3/5
Operational maturity
1/5
Composability
4/5
Local control
5/5
Security boundary clarity
5/5
First-party routing documentation
5/5

Strengths

None documented.

Poor-fit scenarios

  • General SWE
  • No suitable GPU
  • Need cross-machine metric comparability

Limitations

None documented.

prepare.py module
prepare.py

Fixed data/tokenizer/eval utilities.

datatokenizereval
train.py module
train.py

Only file agent modifies.

modeloptimizertraining
program.md configuration
program.md

Human-authored research-agent policy.

instructions
Languages
Python, Markdown
Frameworks
PyTorch
Runtimes
Python >=3.10, uv
Operating systems
None documented
Installation methods
None documented
Interfaces
None documented
Required credentials
None documented
External services
None documented
Hardware
single NVIDIA GPU; tested H100
Major dependencies
None documented

Compatibility notes

None documented.

One-sentence semantic summary

Minimal autonomous LLM research loop: agent edits train.py, runs fixed five-minute experiments, checks val_bpb, and keeps/discards changes.

Capability keywords

agent-edited train loopfixed-time experimentsvalidation metricexperiment log

User intent phrases

  • Need autonomous small-LLM experiments on suitable NVIDIA GPU

Negative match phrases

  • General SWE
  • No suitable GPU
  • Need cross-machine metric comparability

Differentiators

None documented.

Security considerations

  • Run agent with restricted permissions; bound compute/cost.

Privacy considerations

None documented.

Uncertainty

Unresolved items

None documented.

Inference notes

None documented.

Evidence

Repository usage guidance

installation summary
Unknown
basic usage summary
Let a coding agent explore model/optimizer/training changes on a bounded single-GPU setup while the human controls program.md.
documented entry points
None documented
key configuration files
None documented
important directories
None documented
documentation paths
README.md
example paths
None documented

Relationships

downstream
high
Autoresearch uses a simplified nanochat-derived training setup.

Recommended combinations

LLM research loop high

Each repository covers a distinct layer; avoid combining tools that duplicate lifecycle ownership.

Order of operations

  • Understand/reproduce nanochat baseline
  • Set up autoresearch on suitable GPU
  • Define program.md policy
  • Run bounded experiments and review logs

Routing rules

rule_040: Need autonomous small-LLM experiments on suitable NVIDIA GPU P960

Rationale

Minimal autonomous LLM research loop: agent edits train.py, runs fixed five-minute experiments, checks val_bpb, and keeps/discards changes.

Required conditions

None documented.

Preferred

karpathy/autoresearch

Fallback

None