GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
guilindev
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
guilindev
Model Tree
Qwen/Qwen3-0.6B
Input Modalities
Output Modalities
Supported Functionality
This short run is a diagnostic checkpoint, not a successful general game agent. Uniform random legal selection has expected validation agreement 31.67%; always taking the first listed option achieves 38.87%. The LoRA model reaches 32.23%, and selected the second option in all 32 additional inspected states.
A separately trained 17,601-parameter structured baseline reaches 50.20% on the same validation set. Its weights and architecture are released at guilindev/pacman-decision-tiny. This comparison shows why lower training loss alone is insufficient evidence of a useful decision policy. Wider and repeated experiments are still needed.
| Metric on 512 validation states | Original Qwen3 | Fine-tuned adapter |
|---|---|---|
| Top-choice agreement with search teacher | 25.39% | 32.23% |
| Agreement where teacher target > 0.9 (258 states) | 28.29% | 33.33% |
| Soft-target cross-entropy (lower is better) | 3.5232 | 1.1520 |
| Squared distance to teacher distribution (lower is better) | 1.0496 | 0.4256 |
These are recorded measurements from our own run, not results quoted from another
project. Validation means agreement with the simulation teacher; it is not a win
rate. metrics.json contains the exact counts. This first release was shortened
for a live demonstration and uses one training epoch. It is a pilot checkpoint.
Independent gameplay tests use held-out seeds 100–102, realtime clock, speed 1,
and a 180-second simulated-time limit. Results are in gameplay_results.json (all individual seeds included).
| Policy | Mean pellets/game | Mean per-game request p50, rounded ms |
|---|---|---|
| base | 0 | 21 |
| greedy | 173.3 | 0 |
| random | 33 | 0 |
| structured | 304.3 | 1 |
| trained | 49 | 21 |
base and trained are the original and LoRA Qwen models. structured is the
separate tiny policy. Tiny runs on CPU; Qwen runs on GPU, all on the same host.
Zero ms means below the integer-ms reporting resolution, not zero computation.
The comparison has only three games on one fixed maze.
c1899de289a04d12100db370d81485cdf75e47ca, Apache-2.0.The loss is cross-entropy against a soft distribution over legal answer letters. It reuses the language model's output-embedding rows for A/B/C/D. No new dedicated classification head is added. LoRA changes the hidden representation that feeds those option scores. This is related to fixed-choice decision scoring, not a reproduction of Jev's proprietary training method.
We generated new states using grapeot/decision-pacman,
commit 593ed1f59f4f97ad0bda7287ff304e9667360089 (MIT).
An engine search teacher tries each move and rolls forward five seconds with
greedy continuation. Its value scores become a softmax target at temperature 50.
Collector policies include oracle, greedy and random play.
Only current-state features appear in student inputs. Future rollouts and target values are excluded. This is simulation/search-teacher distillation; no 27B teacher, Jev API, private customer data or upstream trained student weights were used. The dataset card explains collection, deduplication, selection and seed ranges.
Download or clone this repository, install a PyTorch build appropriate for your device, then install the remaining requirements. The merged BF16 weight file is about 1.19 GB, plus tokenizer and code. GPU inference uses BF16; the sample script also supports CPU with FP32 weights.
bash
pip install -r requirements.txtpython predict.py --request example.json# After downloading the full model repository:python predict.py --model . --request example.json --device cpuThe script prints a legal move and its candidate distribution. Model loading is
excluded from its printed forward-pass time. It uses the same prompt builder as
training, with enable_thinking=False.
To repeat training from the pinned original base:
bash
python download_data.pypython train.py --limit 4096 --val-limit 512 --batch 16 --epochs 1 --out runs/pilot-v1training-source.zip includes collection and export scripts, the frozen pilot
protocol, and full-run protocol. export_checks.json records adapter merge and
save/reload checks. The release uses safe tensor weights, not Python pickle.
This checkpoint consumes text state features, not game screenshots. It has not been trained as a multimodal model. A visual encoder or multimodal base and paired image/action training data would be a separate experiment.
One fixed maze, synthetic observations, one task. General decision ability and transfer to other games have not been evaluated. Candidate scores are relative to the choices provided; they are not calibrated probabilities of winning or safety. No matched Jev benchmark was run, so this release makes no claim of outperforming Jev.
Qwen3 supplies the base model under Apache-2.0. The game, search oracle, feature
encoder and prompt formatter come from grapeot's MIT-licensed decision-pacman;
see LICENSE.upstream. This project contributes a new data collection, deduplication,
standard Transformers/PEFT training run, export checks and reproducible artifacts.
Merged weights are released under Apache-2.0. This is a personal research project.
pip install -r requirements.txtpython predict.py --request example.json# After downloading the full model repository:python predict.py --model . --request example.json --device cpupython download_data.pypython train.py --limit 4096 --val-limit 512 --batch 16 --epochs 1 --out runs/pilot-v1