win10/Gemma4-HRM-RWKV-7-5B-Token

llamafactory-gemma4-rwkv-hrm-pt-text-freeze

This model is a fine-tuned version of win10/Gemma4-HRM-RWKV-7-Untrained on the gemma4_fineweb_edu_10gb_local and the gemma4_tw_apple_daily_1b datasets.

Model description

Gemma4-HRM-RWKV-7 recovery checkpoint after a text-only PT run with the RWKV/TICA/HRM replacement layers unfrozen.

This is not an instruction-tuned checkpoint. Short text-completion smoke tests produce coherent continuation, while instruction-style greedy generation can still repeat or answer poorly.

Intended uses & limitations

More information needed

Training and evaluation data

Training data used in this run:

• gemma4_fineweb_edu_10gb_local
• gemma4_tw_apple_daily_1b

Total packed training data volume: 5,004,529,664 tokens, or 5.0045B tokens. This is computed from the post-tokenization PT dataset size reported by Trainer, 1,221,809 packed examples × cutoff length 4096.

Tokens actually consumed by this 300-step recovery run: 9,830,400 tokens, or 0.00983B tokens (300 optimizer steps × total train batch size 8 × cutoff length 4096).

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0004
• train_batch_size: 2
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 0.03
• training_steps: 300

Training results

Framework versions

• Transformers 5.6.0
• Pytorch 2.12.0+cu130
• Datasets 4.0.0
• Tokenizers 0.22.2