Hermes-Bonsai Karpathy Self-Improving Agent Loop
Stage 2 checkpoint for the Hermes/Bonsai Karpathy auto-research loop.
Last updated: 2026-04-05
This release is inspired by Andrej Karpathy's framing of self-improving training loops and auto-research. It contains the model artifact that worked, plus a concise model card explaining how it was produced and how to run it.
Overview
- Base model: Qwen3-8B-Base
- Training method: supervised fine-tuning via the Hermes/Karpathy loop
- Stage: Stage 2 — the checkpoint that worked
- Known limitation: Stage 3 exposed a learned-helplessness pattern on some tasks; that behavior is documented in the GitHub methodology repo
- License: Apache-2.0 for this release; the underlying base model license also applies to the inherited Qwen3-8B-Base components
What went into this checkpoint
- The loop-produced training curriculum and trace distillation pipeline
- 140 verified raw passes used as positive reinforcement for curriculum rebalancing and trace selection
- These are Bonsai's own unedited outputs that passed teacher evaluation
- 10 domains covered across the build
- Validation signal from a mixed-domain batch
Domains covered
- memory_integration
- refusal_redirect
- self_correction
- agent_routing
- devops
- logic_puzzle
- code_debugging
- math
- architecture
- research_synthesis
Strongest domains
Best performance concentrated in:
- memory_integration
- refusal_redirect
- self_correction
Validation metrics
- Mixed-domain batch: 13/50 raw passes
- Raw pass rate: 26%
- This checkpoint is the stage 2 model that produced those verified passes
What's novel
Trained via a graduation protocol with teacher-guided validation, raw-pass reinforcement, and frontier failure analysis. The interesting contribution is the loop methodology; see GitHub for the full curriculum and training workflow.
GitHub methodology
The training loop, curriculum design, graduation protocol, and detailed methodology live here:
Files in this Hugging Face repo
bonsai-8b-stage2-post-curriculum-q8.gguf — the shipped stage 2 checkpointREADME.md — this model cardLICENSE — Apache-2.0 license
How to use
Recommended working config from the stable local run:
--ctx-size 40960--n-gpu-layers 37
llama.cpp
1./llama-cli -m bonsai-8b-stage2-post-curriculum-q8.gguf \
2 --ctx-size 40960 \
3 -p "Explain the CAP theorem for a backend engineer."
llama-server
1./llama-server -m bonsai-8b-stage2-post-curriculum-q8.gguf \
2 --ctx-size 40960 \
3 --n-gpu-layers 37 \
4 --host 0.0.0.0 --port 8080
Then point your client at the local OpenAI-compatible endpoint exposed by llama-server.
Notes
- This is a release checkpoint, not the full training corpus.
- The GitHub repo contains the code and documentation needed to reproduce the loop.
- The Hugging Face repo contains the model artifact that ships from that loop.
