cribl-decision-1.0
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity5/10
Apache-2.0 LoRA adapters on Qwen3.5 in four sizes (0.8B, 4B, 9B, and a 27B interim checkpoint). Cards with loading code, the exact prompt layout and a local /v1/systemone example. No hosted API yet; StreamAI gateway support is promised. No SLA.
- Capability6/10
Choice, noul and score read from the lm_head over up to 255 option letters, several questions per pass. Only Cribl's own 26-task held-out numbers; not on JevBench, the Decision Index or Artificial Analysis. No ECE, no latency figures.
- Adoption1/100
About 290 downloads and 5 likes across the four repos; the launch blog post and one X post from Cribl with almost no engagement. No integrations or production users found.
Vendor claims
- 4B on a 26-task held-out set (29,743 records): 77.3% overall accuracy, 78.6% task-macro accuracy, 78.7% macro-F1; beats Jev outright on 6–10 of the 26 tasks[Vendor claim — not independently verified]
- Same held-out set: 9B 78.41% overall; 27B interim checkpoint (step 16,000 of 48,186) 80.73% overall, 82.45% task-macro[Vendor claim — not independently verified]
- Trained on more than 1.5 million records from classification, intent, preference, safety, knowledge, reasoning, agent/tool and rule tasks[Vendor claim — not independently verified]
publicado (OK do Fabricio 19/10 BRT). Scores are calculated from scratch.
Read this first
- We missed the launch. Cribl published the 4B on 9 Oct. It did not show up in our Hugging Face searches until the 9B and 27B repos appeared on 18 Oct.
- The 27B is not finished. Its card calls it an interim checkpoint at about a third of training, with evaluation to follow.
- All accuracy numbers are Cribl’s. None of the sizes is on JevBench, the Decision Index or the Artificial Analysis index (checked 19 Oct).
What it is
cribl-decision is a family of LoRA adapters from Cribl, the telemetry and observability company. You send an input (a log event, a ticket, a request), a question and the options. One forward pass reads the hidden state at the answer slot, projects it through the base model’s own lm_head restricted to the option letters, and returns a softmax over your options. No text is generated.
Cribl’s blog frames it as a general model first: every telemetry event is a routing, enrichment or triage decision made millions of times a day. A smaller model, a telemetry decision benchmark and a domain-tuned version are listed as next steps, and the blog says the model will come to Cribl’s StreamAI gateway “in the coming months”.
| Size | Base | Released | Held-out accuracy (Cribl) |
|---|---|---|---|
| cribl-decision-1.0 (4B) | Qwen3.5-4B-Base | 9 Oct | 77.34% |
| 0.8B | Qwen3.5-0.8B | 9 Oct | 68.52% |
| 9B | Qwen3.5-9B | 18 Oct | 78.41% |
| 27B (interim) | Qwen3.5-27B | 18 Oct | 80.73% |
Specs
| Attribute | Value |
|---|---|
| Vendor | Cribl, Inc. (Hugging Face org cribl-ai) |
| Artifact | LoRA adapters (4B: 57.6 MB; 27B: rank 32, about 95M trainable parameters); base weights frozen |
| Question types | choice (pick + distribution), noul (probability of a yes/no statement), score (ordinal distribution + expected value) |
| Options | Up to 255 per question (A–Z, then two-letter single-token labels); several questions per pass, each read at its own slot |
| Prompt | Fixed layout (Context: / Question: / Options: / Answer: (); the card warns that other formats lose accuracy |
| Serving | Transformers + PEFT with Qwen3_5ForConditionalGeneration (the card warns AutoModelForCausalLM silently ignores the adapter on Transformers ≥ 5.19). The 4B card shows a local POST /v1/systemone call; we found no vLLM, Ollama or gateway listing |
| Context | 8,192 tokens in training |
| License | Apache-2.0 (base models under their own Qwen terms) |
Checked on 19 Oct 2026: 4B 246 downloads and 5 likes; 0.8B 28; 9B 11; 27B 6.
Score working (19 Oct 2026, from scratch, docs/06 formulas)
- Maturity = availability 7 × 0.30 + docs 7 × 0.25 + integrations 2 × 0.25 + support 3 × 0.20 = 2.10 + 1.75 + 0.50 + 0.60 = 4.95 → 5.0.
- Capability = types 10 × 0.30 + calibration 4 × 0.25 + latency 4 × 0.25 + accuracy 5 × 0.20 = 3.00 + 1.00 + 1.00 + 1.00 = 6.00 → 6.0.
- Adoption = engagement 2 × 0.35 + ecosystem 1 × 0.35 + production 0 × 0.30 = 0.70 + 0.35 + 0 = 1.05 → 1.
Availability 7: open weights with a clear licence, no hosted API (Jeva-4B precedent). Docs 7: detailed cards with loading code, readout code, prompt layout and a runnable script, plus the blog. Integrations 2: a local /v1/systemone example in the card, StreamAI support only promised. Support 3: company-backed, no SLA for the model. Types 10. Calibration 4: tagged “calibrated”, no ECE published, and the card says to recheck calibration on your own data. Latency 4: no figure published. Accuracy 5: held-out sets built by Cribl from public data, no public board. Engagement 2, ecosystem 1, production 0.
Fit / anti-fit
Fit: self-hosted routing, enrichment and triage of log or security events where you control the prompt format and can test on your own data; teams already on Qwen3.5 and PEFT.
Anti-fit: anything that needs a hosted endpoint or a TypeSafe client today, the 27B before its final checkpoint, and any choice based on Cribl’s numbers alone. The card lists overconfidence and sensitivity to option order and wording among its limits.