NanoJev-1.7B
A 1.7B variant of C-Tianyu/NanoJev using Qwen3-1.7B as the backbone, with approximately 1.7B parameters.
It retains the structured decision modeling approach of NanoJev while scaling the backbone from 0.6B to 1.7B parameters. The model produces complete probability distributions over dynamic decision candidates without relying on output-token decoding.
Model
| Property | Value |
|---|---|
| Backbone | Qwen3-1.7B |
| Parameters | ~1.7B |
| Model weights | BF16 |
| Training optimizer | 8-bit AdamW |
Note: The model weights are BF16. The "8-bit" designation refers to the AdamW optimizer used during training, not to INT8/GPTQ/AWQ weight quantization.
Accuracy
On a manually constructed 17-case evaluation set:
| Model | Correct | Accuracy |
|---|---|---|
| Original NanoJev | 16 / 17 | 94.12% |
| NanoJev-1.7B | 17 / 17 | 100.00% |
Decision Confidence
| Model | Average Selected-Decision Probability |
|---|---|
| Original NanoJev | 53.58% |
| NanoJev-1.7B | 86.97% |
| Improvement | +33.39 percentage points |
Test Metrics
| Metric | NanoJev-1.7B |
|---|---|
| Target CE | 0.6077 |
| Target KL | 0.3691 |
| Target TV | 0.2313 |
Usage
The model can be used with the NanoJev inference and decision-prediction pipeline.
The model outputs probability distributions over structured decision candidates rather than generating decision labels as ordinary text.
Notes
The 17-case accuracy evaluation is a small, manually constructed evaluation set and should not be considered a standardized benchmark.
The reported results are based on the author's evaluation data and may not represent general real-world performance.
License
This model uses Qwen3-1.7B as its backbone and is distributed under the Apache License 2.0.
The NanoJev project is licensed under the MIT License. Please refer to the respective original licenses for applicable terms and conditions.
- Downloads last month
- 17