Downloads · 30 days
17
34% of all-time downloads
steven0226/qwen2.5-1.5b-wordle-grpo
qwen2.5-1.5b-wordle-grpo is a text generation model from steven0226. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as apache-2.0.
This repository contains a LoRA adapter, not standalone full-model weights. It was trained from the base repository identity Qwen/Qwen2.5-1.5B-Instruct for a multi-turn Wordle environment using GRPO. The adapter must…
Downloads · 30 days
17
34% of all-time downloads
All-time downloads
50
Public
Repo size
85.3 MB
Likes
0
Public
Click a slice to open those files.
.safetensors73.9 MB · 87%
From the Hugging Face model README
This repository contains a LoRA adapter, not standalone full-model weights. It was trained
from the base repository identity Qwen/Qwen2.5-1.5B-Instruct for a multi-turn Wordle environment
using GRPO. The adapter must be loaded together with the compatible base model.
The PEFT configuration uses rank 16, alpha 32, dropout 0.05, and causal language modeling as the
task type. The current adapter_model.safetensors LFS SHA-256 is
92e6379ed7ddf363e7f500b143afa7a2dc725d3e86bd87bc9eb933831c7d68b7.
The reported tuned evaluation is the adapter evaluation. It is a paired, greedy evaluation over the complete held-out 463-word split with seed 42. Only committed aggregate evidence is available.
| Measure | Base | Tuned LoRA adapter |
|---|---|---|
| Wins | 0/463 | 13/463 = 2.81% |
| Wilson 95% CI | 0.00%–0.82% | 1.65%–4.74% |
| Protocol adherence | 0% | 2749/2753 = 99.85% |
| Legal actions | 0% | 2748/2753 = 99.82% |
The paired discordance was 0 base-only wins and 13 tuned-only wins. The two-sided exact paired
McNemar result is 0.000244140625; treating the interim and final evaluations as two looks gives
the conservative Bonferroni-adjusted value 0.00048828125.
Strategy use remained weak: absent letters were reused on 1340/2290 turns with information, and known green positions were broken on 1119/2290 such turns.
Protocol learning succeeded; strategy learning remained limited; this is not a practical Wordle solver. The 2.81% win rate is statistically distinguishable from this base evaluation, but it is not evidence of practical solving capability or general RL superiority.
results/full_463_report.jsonresults/full_463_analysis.jsondocs/claim-matrix.mdLoad the compatible base model first, then attach this LoRA adapter with PEFT. Because the exact historical upstream base revision is unavailable, do not present a contemporary Qwen revision as the cryptographically proven training revision.
from peft import PeftModel
from transformers import AutoModelForCausalLM
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
model = PeftModel.from_pretrained(base, "steven0226/qwen2.5-1.5b-wordle-grpo")