santos-sanz/spanish-wordle-lfm2.5-2.6b-qlora

Spanish Wordle LFM2.5 2.6B QLoRA

Task-specific MLX LoRA adapter trained to play five-letter Spanish Wordle in Pure, Agent, and Oracle modes. The adapter does not contain the 2.6B base weights and must be loaded together with LiquidAI/LFM2.5-2.6B-MLX-6bit at revision 95f71f1c30e3247bc7f042c6fd64d7ca60258780.

Training

• —Base checkpoint: 6-bit MLX export of LiquidAI LFM2.5 2.6B
• —LoRA: rank 16, scale 32, dropout 0.05, final 16 layers
• —Optimizer: AdamW, learning rate 3e-5 for the main SFT phase
• —Batch size: 1 with gradient accumulation 8
• —Maximum sequence length: 512
• —Seed: 20260814
• —Dataset modes: approximately 60% Pure, 30% Agent, 10% Oracle
• —Training targets: 431 of 616 unique historical Spanish Wordle answers
• —Validation targets: 61; hidden test targets: 124

The main SFT run used 5,000 optimizer iterations with gradient accumulation of

1. 1.A subsequent Pure-only cleanup phase used 150 iterations at 1e-5 on 8,757 training and 1,261 validation records. Repair examples keep the invalid guess as runtime context and supervise only the valid replacement, avoiding a repeated-guess label in the loss. The selected local adapter is the final checkpoint of that cleanup phase (global iteration 5,355).

The word lists and reproducible training/evaluation code are maintained in the private repository santos-sanz/spanish-wordle-slm. Test targets were excluded from training trajectory generation.

Evaluation status

The frozen comparison uses deepseek/deepseek-v4-flash-0731 through OpenRouter, temperature 0, seed 20260814, reasoning low, no provider fallback, and a 512-token cap. Pure has no tools; Agent may call get_candidates once per turn; Oracle is a solver ceiling and does not count toward the competitive claim. The final paired bootstrap report in the private repository is authoritative. Success requires a positive 95% interval for the SLM in both Pure and Agent; otherwise the run is reported as an experiment that did not meet the target.

The hidden test produced 2/124 wins for the SLM versus 1/124 for the benchmark in Pure (paired 95% CI for the win-rate difference: [-1.6%, +4.0%]), and 115/124 versus 8/124 in Agent (CI: [+79.0%, +91.9%]). The adapter therefore demonstrates a statistically clear Agent-track win, but this run does not meet the two-track success claim because Pure is inconclusive.

Use with MLX-LM

python -m mlx_lm.generate \
  --model LiquidAI/LFM2.5-2.6B-MLX-6bit \
  --adapter-path santos-sanz/spanish-wordle-lfm2.5-2.6b-qlora \
  --prompt 'Juega Wordle en español.'

License and attribution

This adapter is a derivative of LiquidAI LFM2.5 and is distributed under the LFM 1.0 License. Review that license before use or redistribution. LiquidAI is the author of the base model; this project only provides task-specific adapter weights.