emender-e97-1.3b-cli-agent

Emender E97 1.3B CLI Agent

A narrow experimental recurrent CLI policy derived from the 513B-token dense E97 base. It is trained for one exact RS-free Pi protocol:

repository objective -> sandboxed direct CLI argv -> typed observation -> grounded submit_answer

This is not a general instruction or coding assistant.

Verified behavior

The immutable checkpoint passed 40/40 held-out real-Pi direct CLI tasks:

All 40 runs used exactly cli -> submit_answer and passed argv, value, evidence, grounding, bounded-sequence, completion, and clean-exit checks.

• Source checkpoint SHA-256: fca9b06c521fb9407cd7bed2d7049f36e9a93948d6cd98899b6ad9b9e4fc6b01
• Base: dense E97 513B authority
• Weight semantics: saved evaluated agent weights
• Tokenizer: p50k_base
• Raw pickle checkpoint: not included

Important protocol dependency

The result depends on the exact system prompt, tool schema, assistant-target normalization, structured-action stopping, recurrent cache semantics, and hash-pinned cwd-only Apptainer sandbox in the Emender repository. Ordinary text generation does not reproduce the agent evaluation.

See configs/pi/e97-cli-tools.ts, scripts/run_e97_sandbox_cli.py, and docs/E97_DENSE_PI_AGENT_EXECUTION_PLAN.md at commit 2b5d135587cca9d1081dbbdf7b35a8ac2629db5d.

Loading weights

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("spinozans/emender-e97-1.3b-cli-agent")
model = AutoModelForCausalLM.from_pretrained(
    "spinozans/emender-e97-1.3b-cli-agent", trust_remote_code=True, dtype="auto"
)

Install the matching model implementation:

pip install "git+https://github.com/spinozans/emender.git@2b5d135587cca9d1081dbbdf7b35a8ac2629db5d"

Limitations and license

The policy has not demonstrated autonomous tool discovery or general software engineering ability. Its sandbox does not claim arbitrary hostile-code isolation. No standalone model license has yet been selected; license: other is intentional.