--- license: mit tags: - constrained-decoding - embodied-agent-interface - llama-3 - gpt-oss pretty_name: Meta-Ctrl EAI artifacts (Llama 3 8B, gpt-oss-20b) --- # Meta-Ctrl: EAI artifacts Artifacts for the Embodied Agent Interface experiments of Meta-Ctrl with two base LMs, Llama 3 8B Instruct and gpt-oss-20b. The code repository (https://github.com/yidouweng/meta-ctrl) fetches and unpacks them with `scripts/download.sh`. | Archive | Contents | Unpacks to | |---|---|---| | `outputs.tar.gz` | generations and EAI evaluator outputs: Meta-Ctrl with both LMs, the Llama ablation arms, the search, classical-planner and XGrammar baselines, the lookahead calibration | `outputs/`, `results/` | | `models.tar.gz` | the meta-token HMMs for both LMs, trained and uniform (`models/README.md`) | `models/` | | `train_data/llama_samples.tar` | the Llama 3 8B Instruct samples the Llama HMMs were trained on | `work/samples/` | | `train_data/gpt-oss_samples.tar` | the gpt-oss-20b samples (final channel) the gpt-oss HMMs were trained on | `work/samples/` | | `train_data/llama_corpus_as.tar`, `train_data/llama_corpus_sd.tar` | the Llama meta-token training corpora, including the balanced subsets | `work/corpora/` | `python scripts/tables.py --check` in the code repository prints every table from `outputs.tar.gz` and compares it with the expected values.