Downloads · 30 days
0
santos-sanz/spanish-wordle-lfm2.5-2.6b-qlora
spanish-wordle-lfm2.5-2.6b-qlora is a text generation model from santos-sanz. Use it when you need the model to write or continue text. It is set up for mlx. The card lists the license as other.
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-…
Downloads · 30 days
0
Access
Public
Updated Aug 16, 2026
Repo size
65.2 MB
Likes
0
Public
Click a slice to open those files.
.safetensors52.2 MB · 100%
From the Hugging Face model README
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.
3e-5 for the main SFT phase20260814The main SFT run used 5,000 optimizer iterations with gradient accumulation of
8. 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.
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.
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.'
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.