HRM-Grammar-Light

Overview

• Task: Causal grammar correction given a prompt "corregir {idioma}: ".
• Architecture: RMSNorm, SwiGLU FFN, multi-head attention with RoPE, separate H/L stacks, pooled ACT halting (Graves-style), optional deep supervision, tied embeddings.
• Trainer: nnm.py implements data prep, training, validation, and Hub uploads.
• Hub model: dreuxx26/HRM-Grammar-Light (weights/checkpoints uploaded by the trainer).

Data

• input_text: Original text to correct. Optionally prefixed like corregir español: ... (if not present, the script infers/cleans it).
• target_text: Corrected text target.
• Optional language: If present, it’s normalized; rare classes are collapsed to other for stable splits.

1. From Hugging Face Datasets Hub (recommended for contributors)

• Set env HRM_DATASET_HUB_ID (e.g., your_user/hrm_grammar_light). Optional: HRM_DATASET_CONFIG, HRM_DATASET_SPLIT, HRM_DATASET_REVISION.

1. From Google Drive (legacy/local)

• Place Arrow files under `/content/drive/MyDrive/my_dataset/ . dataset: https://huggingface.co/datasets/dreuxx26/idiomas_5

Training (Colab-friendly)

• Secrets: Save HF_TOKEN in Colab Secrets to enable Hub uploads.
• Run nnm.py in Colab (the script auto-mounts Drive and installs deps when in Colab).
• The trainer builds sequences as: "corregir {idioma}: " + </s> + target + </s> and masks the prompt (labels = -100) so loss is on the target only.

• Mixed precision autocast: bf16 on A100 (Ampere+), fp16 on T4 automatically.
• TF32 enabled only on Ampere+.
• Gradient checkpointing (toggle with HRM_CHECKPOINTING).
• Token-weighted epoch averages and validation metrics (loss, token-acc, normalized edit distance). Jump in and train
  Colab (1-click): https://colab.research.google.com/drive/1pMrC_Mgk-EcULxvYCj6QsfLic0VlBxGF?usp=sharing
  original: https://colab.research.google.com/drive/1Q9pqGU8pP79qYX0josznBZzo6HrQsOYA?usp=sharing

Community Mode (safe contributions)

• Set HRM_COMMUNITY_MODE=1 and HRM_CONTRIB=<handle>.
• The script writes a manifest.json and uploads to contrib/<handle>-<timestamp>.
• Upload gating: by default, only uploads when validation loss improves. To allow uploads even without improvement, set HRM_ALLOW_REGRESSION=1.
• Commit message includes BEST when it’s the best val loss so far.

Important environment variables

• Project
  • HF_REPO_ID: Model repo to upload (default: dreuxx26/HRM-Grammar-Light).
• Data
  • HRM_DATASET_HUB_ID, HRM_DATASET_CONFIG, HRM_DATASET_SPLIT, HRM_DATASET_REVISION.
  • LOCAL_DATASET_PATH: Drive folder with .arrow files (fallback when no Hub dataset).
• Training
  • HRM_EPOCHS (default 10), HRM_BLOCK_SIZE (default 256), HRM_BATCH_SIZE (default 8), HRM_LR (default 3e-4), HRM_WARMUP_RATIO (default 0.1).
  • HRM_EFF_BATCH: target effective batch size (auto-adjusts grad accumulation).
  • HRM_CHECKPOINTING (1/0), HRM_ACT_STEPS (default 8), HRM_ACT_EPS (default 0.01).
  • HRM_DEEP_SUPERVISION (0/1), HRM_DS_WEIGHT (0.2 by default).
• Community/Uploads
  • HRM_COMMUNITY_MODE (0/1), HRM_CONTRIB, HRM_UPLOAD_BRANCH (optional explicit branch), HRM_ALLOW_REGRESSION (0/1).

T4 tips

• Keep HRM_BLOCK_SIZE at 192–256.
• Use HRM_EFF_BATCH to reach your desired effective batch while avoiding OOM.
• Leave checkpointing on (HRM_CHECKPOINTING=1).
• Mixed precision is automatic (fp16 on T4).

How uploads work

• If HF_TOKEN is present, the trainer saves weights (pytorch_model.bin), a checkpoint state (local_training_state.pt), and manifest.json locally.
• If in Community Mode, it uploads to contrib/<handle>-<timestamp>; otherwise to main.
• Uploads are gated by validation improvement unless HRM_ALLOW_REGRESSION=1.

Troubleshooting

• OOM on T4: lower HRM_BLOCK_SIZE, increase HRM_GRAD_ACCUM via HRM_EFF_BATCH, keep checkpointing on.
• HF auth errors: set HF_TOKEN in Colab Secrets or as an environment variable.
• Dataset errors: verify HRM_DATASET_HUB_ID exists and is public (or authenticated), or ensure Arrow files are in the Drive path.

License & contributions

• This project is licensed under Apache-2.0. You’re free to use, modify, and distribute—with attribution and notice.
• Contributions are welcome via Community Mode PRs/branches on the Hub.