Model Experiment Workflow
Structured trial-based workflow for Qwen3-ASR ONNX model optimization. Each experiment modifies one variable, measures WER/RTF, and records results in INVESTIGATION.md regardless of outcome.
Workflow
1. Read prior work
Read the top of INVESTIGATION.md to find the latest experiment number. The next experiment is [N+1]. Scan recent experiments for context on what's been tried and current baselines.
Current baselines (update from INVESTIGATION.md if stale):
- 0.6B int4 RTN al4: ~5.16% WER, ~0.17x RTF
- 1.7B int4 RTN al4: ~4.20% WER, ~0.29x RTF
2. Set up trial directory
Create a trial directory with symlinks to avoid duplicating multi-GB files. Only the file under test should be a real copy or new file.
# Example: testing a new encoder variant
cp -rs $(pwd)/output/qwen3-asr-0.6b output/trial-<name>
rm output/trial-<name>/encoder.int4.onnx # remove symlink for the file under test
cp <new-encoder> output/trial-<name>/encoder.int4.onnx
The trial dir must contain ALL files needed by evaluate_wer.py including tokenizer_config.json, added_tokens.json, vocab.json — missing any causes AutoTokenizer.from_pretrained() to fall through to AutoConfig which fails on model_type: qwen3_asr.
3. Run WER evaluation
cd ~/qwen3-asr-onnx
uv run python evaluate_wer.py \
--models \
"baseline:output/qwen3-asr-0.6b:int4" \
"trial-name:output/trial-<name>:int4" \
--datasets librispeech-other \
--n-samples 200 \
--output /tmp/wer_<experiment>.json
The :int4 suffix resolves encoder.int4.onnx, decoder_init.int4.onnx, etc. Omit for FP32. Run in background for long evals (~30-60 min for 200 samples with 2 models).
Check process is alive: ps aux | grep evaluate_wer
Output is buffered — check file size growth: wc -c <output-file>
Final results appear in the last lines with a summary table.
4. Record results in INVESTIGATION.md
Prepend the new experiment at the TOP of the file (newest first). Use this format:
### [N] Title — brief description
**Date:** YYYY-MM-DD
**Idea:** What hypothesis is being tested and why. Reference prior experiments.
**Change:** What was modified (files, scripts, parameters).
**Result (200-sample LibriSpeech test-other):**
| Trial | Description | WER | RTF | Size |
|---|---|---|---|---|
| Baseline | ... | X.XX% | X.XXx | X MB |
| Trial | ... | X.XX% | X.XXx | X MB |
**Outcome:** SUCCESS/FAILURE/DEGRADED/FINDING — one-line summary.
**Notes:**
- Why it worked or failed (root cause analysis)
- Implications for future experiments
- References to related experiments [N]
Outcome categories:
- SUCCESS: meets or exceeds target, ready for adoption
- FAILURE: broken output, crashes, or catastrophic degradation
- DEGRADED: works but measurably worse than baseline (document the tradeoff)
- FINDING: neutral result that informs future work (e.g. measurement, characterization)
5. Commit or revert
On SUCCESS:
- Commit model file changes (new scripts, quantized files, config updates)
- Commit INVESTIGATION.md entry
- Update README.md if the finding changes recommended variants or package contents
- Update memory with key findings
On FAILURE/DEGRADED:
- Revert any model file changes:
rm -rf output/trial-<name> - Commit ONLY the INVESTIGATION.md entry — negative results are valuable
- Update memory if the finding is surprising or corrects assumptions
6. Clean up
rm -rf output/trial-<name>
rm /tmp/wer_<experiment>.json
Disk space constraints
WSL VHD grows but never shrinks. Check C:\ before large operations: df -h /mnt/c. Use symlinks/hardlinks for trial dirs. Delete trial dirs promptly after recording results.
Tool reference
See references/tools.md for parameters and usage of all Python tools in the repo.
Source: andrewleech/qwen3-asr-onnx — distributed by TomeVault.