run_2026-09-22_qwen3
v3.4 · Qwen3-235B comparison
Qwen3-235B · bge-m3 · hybrid (BM25 + dense, RRF) · support-prompt v7 on support-golden-v4 v4.6
Pass rate
80.2%
Faithfulness
0.87
Answer correctness
0.81
Context precision
0.79
Context recall
0.84
Latency p95
4.4s
Failures by category
246 failed samples · largest share: unsupported
| Category | Share | vs reference |
|---|---|---|
| Unsupported claims | 26.0% | −1.1 pts |
| Incorrect factual answers | 22.0% | +1.0 pts |
| Missing context | 15.9% | −3.3 pts |
| Retrieval mismatch | 15.0% | −2.9 pts |
| Instruction-following failures | 15.0% | +3.9 pts |
| Other | 6.1% | +2.4 pts |
Run details
- Dataset
- support-golden-v4 v4.6 · 1,240 items
- Generator
- Qwen3-235B
- Retriever
- bge-m3 · hybrid (BM25 + dense, RRF)
- Prompt
- support-prompt v7
- Judge
- LLM judge · Claude Opus 5.5 · rubric v3
- Started
- Sep 22, 10:30 UTC
- Duration
- 3h 21m
- Triggered by
- Sara Lindqvist
- Tokens / request
- 3,420 in · 228 out
- Compared with
- support-assistant v3.4 · release candidate
- open-weights
- comparison
All metrics
Against support-assistant v3.4 · release candidate
| Metric | Group | This run | Reference | Change |
|---|---|---|---|---|
| Pass rateShare of samples the judge marked as passing all rubric checks. | Outcome | 80.2% | 86.9% | −6.7 pts |
| FaithfulnessShare of answer claims supported by the retrieved context. | Generation | 0.87 | 0.91 | −0.04 |
| Answer correctnessAgreement with the gold answer on facts, not style. | Generation | 0.81 | 0.88 | −0.07 |
| Instruction adherenceRespect for tone, format, length, language and policy constraints. | Generation | 0.86 | 0.93 | −0.07 |
| Context precisionRank-weighted precision of relevant chunks passed to the generator. | Retrieval | 0.79 | 0.79 | 0.00 |
| Context recallShare of gold-answer statements supported by the retrieved context. | Retrieval | 0.84 | 0.84 | 0.00 |
| Recall@5Share of gold passages found in the top 5 retrieved chunks. | Retrieval | 0.88 | 0.88 | 0.00 |
| MRRMean reciprocal rank of the first relevant chunk. | Retrieval | 0.74 | 0.74 | 0.00 |
| NDCG@10Graded ranking quality of the top 10, normalized by the ideal order. | Retrieval | 0.81 | 0.81 | 0.00 |
| Latency p50Median end-to-end time to final token. | Operational | 2.0s | 1.9s | +0.1s |
| Latency p9595th percentile end-to-end time to final token. | Operational | 4.4s | 3.8s | +0.6s |
| Cost per requestModel, embedding and reranking cost per answered request (USD). | Operational | $0.0017 | $0.0042 | −0.0025 |
Nightly trend
Nightly runs on the support-golden-v4 slice, Jul 11 to Oct 8. Dashed line marks this run.
- Pass rate · nightly
- This run · 80.2%
Across datasets
v3.4 · Qwen3-235B on every dataset in the project. Rows marked “matrix” come from the comparison matrix rather than a dedicated run.
| Dataset | Items | Pass rate | Faith. | Ctx recall | Recall@5 | p95 | Source |
|---|---|---|---|---|---|---|---|
| support-golden-v4this run | 1,240 | 80.2% | 0.87 | 0.84 | 0.88 | 4.4s | This run |
| billing-edge-cases | 312 | Not evaluated with this configuration | |||||
| roaming-and-travel | 188 | Not evaluated with this configuration | |||||
| device-troubleshooting | 426 | Not evaluated with this configuration | |||||
| account-security | 154 | Not evaluated with this configuration | |||||
| locale-formatting | 96 | Not evaluated with this configuration | |||||
| synthetic-hard-negatives | 540 | Not evaluated with this configuration | |||||
Traced samples
17 samples with full traces (4 failing, 13 passing), stratified by domain. Open a row to inspect retrieval and the judge's reasoning.
What is the excess on a theft insurance claim? Returns & warranty Passjudge 0.90 |
My phone shows 4G but mobile data does not work after the iOS update. Device troubleshooting Passjudge 0.86 |
How do I get a copy of the personal data Harbor holds about me? Data privacy requests Passjudge 0.93 |
What is the late fee if I pay my bill after the due date? Billing & payments Passjudge 0.95 |
How long does it take to port my number to Harbor? Number porting Passjudge 0.96 |
Does the Fall 2026 upgrade offer work with a family plan discount? Promotions Passjudge 0.88 |
If I upgrade my plan mid-month, how is my next bill calculated? Plans & upgrades Passjudge 0.90 |
How much data can I use in the EU on the Unlimited Plus plan? Roaming & travel Passjudge 0.95 |
Why was I charged twice for my September bill? Billing & payments Passjudge 0.95 |
Can I return a phone I bought online if I opened the box? Returns & warranty Passjudge 0.92 |