Qwen3-1.7B DAPO block-layer16 ReProbe
This repository contains a step-level uncertainty head for
Qwen/Qwen3-1.7B. It scores block-segmented reasoning steps from decoder layer
16. Lower sigmoid probabilities indicate lower predicted step error risk.
The published weights are the pre-selected epoch-2 checkpoint from training
seed 43. Seed 43 is the median-performing block_layer16 run across seeds 42,
43, and 44, and was selected for publication to avoid cherry-picking the
highest dev seed.
Files
weights.pth: PyTorch state dictionary for the uncertainty head.config.yaml: loader-compatible head architecture and feature-extractor
configuration.metadata.json: machine-readable training, metric, and provenance contract.SHA256SUMS: package digests.
This is not a standalone language model. It requires the pinned Qwen3-1.7B base
model and the uncertainty-head implementation from
uncertainty4reasoning.
Base model and architecture
| Field | Value |
|---|
| Base model | Qwen/Qwen3-1.7B |
| Base revision | 70d244cc86ccca08cf5af4e1e306ecf908b1ad5e |
| Segmentation | Block reasoning steps |
| Hidden-state layer | Decoder layer 16 of 28 |
| Head type | step_reasoning |
| Head dimension | 768 |
| Head layers / attention heads | 2 / 8 |
| Dropout | 0.1 |
| Future-token masking | Enabled |
| Training seed / selected epoch | 43 / 2 |
Loading
1from transformers import AutoModelForCausalLM
2from luh import AutoUncertaintyHead
3
4base_model = AutoModelForCausalLM.from_pretrained(
5 "Qwen/Qwen3-1.7B",
6 revision="70d244cc86ccca08cf5af4e1e306ecf908b1ad5e",
7 torch_dtype="auto",
8)
9uhead = AutoUncertaintyHead.from_pretrained(
10 "JingweiNi/ReProbe-Qwen3-1.7B-DAPO-block-layer16",
11 base_model,
12)
13uhead.eval()
Correct inference also requires the training-compatible chat template, stored
token IDs, block-step boundaries, completion offsets, and claim masks. See
bestofn_optimized/score_qwen3_reprobe_3arm.py in the linked code repository
for the exact fused-scoring contract. Do not retokenize a trace and assume its
step masks still align.
Direct checkpoint metrics
The checkpoint was selected by external-dev PR-AUC with early-stopping patience
3. Dev labels are a fixed synthetic-proxy set produced by a Qwen3.6-27B judge.
| Metric | Seed 43, epoch 2 |
|---|
| PR-AUC | 0.410679 |
| PR-AUC / prevalence | 4.046096 |
| ROC-AUC | 0.799337 |
| Thresholded F1 | 0.310571 |
Across seeds 42, 43, and 44, block_layer16 obtained PR-AUC
0.408683 +/- 0.007027 (sample standard deviation). Layer 16 had a higher
block-step dev PR-AUC than the last layer in all three seeds, but every
per-seed question-bootstrap interval included zero. This is consistent
directional evidence, not statistically decisive superiority.
Training data contract
| Split | Finite step labels | Positive labels | Positive rate |
|---|
| Train | 10,000 | 1,320 | 13.20% |
| Dev | 2,000 | 203 | 10.15% |
Train and dev question/source fingerprints were disjoint. The labels are
synthetic step-correctness judgments, not human annotations or native DAPO
ground truth.
Best-of-8 evidence and aggregation
The uploaded seed-43 weights were not the checkpoint used in the downstream
Best-of-8 experiment. That frozen test used the same recipe's seed-42 epoch-2
checkpoint:
- primary pre-registered
max aggregation: 71.00% (213/300);
- shortest-response baseline: 70.33% (211/300);
- exact random-selection expectation: 66.08%;
- layer16 minus shortest: +0.67 percentage points, 95% paired bootstrap interval
[-2.33, +3.67].
After opening the test result, a post-hoc mean reduction reached 73.33%
(220/300) with the seed-42 checkpoint. Its gain over seed-42 max was +2.33
points with interval [-0.33, +5.33]. This reduction comparison is exploratory
and must not be treated as a validated setting for the uploaded seed-43 head.
Limitations
- The checkpoint is tied to Qwen3-1.7B at the pinned revision, decoder layer 16,
and block segmentation.
- Training/dev supervision is synthetic and judge-dependent.
- The 300-question Best-of-8 test set has been consumed and must not be reused
for checkpoint or aggregation selection.
- The experiment does not establish that layer 16 generally beats the last
layer or that ReProbe reliably beats shortest-response selection.
- Calibrate or validate the trajectory aggregation on fresh data before using
the head for selection in another domain.
Integrity
The published weights.pth must have SHA-256
29ebb9df957beae16a3440f4466be22ac8e451d26282af5e738c4a2b6e48b512.
The package's SHA256SUMS file covers the weights, config, model card, and
metadata.
License
Apache-2.0. The base model is also published under Apache-2.0; users remain
responsible for complying with the base model's terms and applicable law.
