soufianeboulahcen/Multimodal-Moroccan-SLG

multimodal moroccan sign language generation # Multimodal Moroccan Sign Language Generation Strict paper-faithful PyTorch reimplementation of **SignLLM** (Fang et al., 2024) trained on the **Moroccan Sign Language (MoSL) video dataset** (Ben Zaid et al., 2026). > **Headline finding.** On the MoSL isolated-word dataset, our strict > paper-faithful SignLLM-Base reimplementation fails to outperform > deterministic retrieval baselines. All three loss configurations from the > paper's Table 5 ablation (MSE / RL / RL+PLC) underperform Nearest-Neighbor > (test DTW 0.78), Mean-Pose (0.87), and Random-Clip (0.96), with our best > trained model at 1.04. See `docs/RESULTS.md` for the > full result table, ablation analysis, and the regression-to-the-mean > diagnostic that explains the gap. --- ## Repository layout ``` . ├── mosl/ Python package (model, data, text, pose, train) ├── scripts/ CLI entry points (training, evaluation, figures) ├── patches/ Our patches to upstream third-party code ├── data/ Dataset metadata (CSVs) and tokenizer vocab ├── docker/ Container build (NGC PyTorch 26.04 base) ├── docs/ Methodology + decisions + results + walkthroughs ├── predictions/ Sample model outputs (NPZ + animated GIF) └── runs/ Tracked: evaluation.json + baselines.json (training logs and checkpoints not redistributed) ``` Each `docs/*.md` file is the source of truth for one aspect of the project: | File | Contents | |---|---| | `docs/STATS.md` | Dataset statistics: counts, long-tail, FPS, splits | | `docs/PIPELINE.md` | End-to-end preprocessing pipeline | | `docs/POSE_EXTRACTION.md` | Phase 2 completion summary | | `docs/PROMPT2SIGN.md` | What we adapted from upstream | | `docs/MODEL.md` | Architecture spec + open questions | | `docs/DECISIONS.md` | Every non-trivial decision, dated | | `docs/RESULTS.md` | Final results, baseline comparison, analysis | | `docs/ …