GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Cochon123
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Cochon123
Model Tree
Qwen/Qwen3-4B-Instruct-2507
Input Modalities
Output Modalities
Supported Functionality
All rows use greedy decoding, the repository's exact prompt and strict move validator, and up to three validator-feedback attempts. The denominator is the same 158 points used by the five-board leaderboard runs.
| Stage | Points | Score | Exact optimal | Legal |
|---|---|---|---|---|
| Base Qwen3-4B | 3/158 | 1.90% | 0% | 20% |
| Clean QLoRA, benchmark hashes excluded | 0/158 | 0% | 0% | 0% |
| Broad 100-board specialization | 40/158 | 25.32% | 20% | 20% |
| Focused five-board specialization (this adapter) | 158/158 | 100% | 100% | 100% |
The final run took 51.4 seconds on a Colab T4. On the 2026-07-31 leaderboard
snapshot this beats Codex CLI gpt-5.6-sol xhigh (83/158, 52.53%) and ties a
cursor/Grok entry at the benchmark's mathematical maximum (158/158).
The exact JSON evaluations are included in this repository.
Training proceeded in three stages:
All stages used a 4-bit NF4 base, float16 compute, LoRA rank 32, alpha 64, dropout 0.05, all linear layers, completion-only cross-entropy, and seed 3407.
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerbase_id = "Qwen/Qwen3-4B-Instruct-2507"adapter_id = "Cochon123/Qwen3-4B-Scrabble-Smoke-Specialized"tokenizer = AutoTokenizer.from_pretrained(adapter_id)base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")model = PeftModel.from_pretrained(base, adapter_id)
Use scrabble_bench.runner.prompt_for_position from the benchmark repository
to construct the exact messages.
Use this adapter to reproduce the public smoke-set ceiling and study fine-tuning dynamics. Do not cite its 100% score as evidence of general Scrabble skill. The clean adapter's official smoke result was 0%, and the broad specialization stage showed clear held-out loss regression.