🚨 GLiNER returns uniform low scores with transformers v5.0.0 (released today)

[elvismdev]

Summary

Transformers v5.0.0 was released 3 hours ago and causes GLiNER to return uniform low confidence scores (~0.05) for all entity types, effectively breaking entity extraction.

Environment

• GLiNER: 0.2.23
• Transformers: 5.0.0 (broken) / 4.57.3 (working)
• Platform: Jetson AGX Orin (ARM64) with CUDA 12.6
• Base image: dustynv/pytorch:2.6-r36.4.0-cu128
• Model: EmergentMethods/gliner_large_news-v2.1

Symptoms

1. Model loads successfully without errors
2. predict_entities() returns entities but with near-identical low scores
3. All tokens receive ~0.05 confidence regardless of actual entity type
4. The model always selects the last label in the label list

Debug output example:

Text: "Donald Trump met with Nicolas Maduro in Cuba"
Labels: ["person", "organization", "location"]

Result with transformers 5.0.0:
- "Cuba" -> location (0.051)
- "Trump" -> location (0.049)  # Wrong! Should be person
- "Maduro" -> location (0.048) # Wrong! Should be person

Result with transformers 4.57.3:
- "Donald Trump" -> person (0.86) ✓
- "Nicolas Maduro" -> person (0.86) ✓
- "Cuba" -> location (0.90) ✓

Root Cause

Transformers v5.0.0 includes a breaking change to default dtype handling:

The library now loads models in their original saved dtype rather than forcing float32.

This likely affects GLiNER's span scoring mechanism, causing the uniform low scores.

Workaround

Pin transformers to v4.x in your dependencies:

transformers>=4.57.3,<5.0.0

Dockerfile example:

RUN pip3 install --no-cache-dir \
    "transformers>=4.57.3,<5.0.0" \
    gliner>=0.2.23

Suggestion

Consider:

1. Adding an upper bound to transformers in setup.py/pyproject.toml until compatibility is confirmed
2. Testing against transformers v5.0.0 to identify the specific breaking change
3. Adding a runtime warning if transformers >= 5.0.0 is detected

Related

• Transformers v5.0.0 Release Notes
• Possibly related to #306 which addressed earlier dtype issues

---

Happy to provide more debug info if helpful. Thanks for the great library! 🙏

[Ingvarstep]

@elvismdev , thank you for being so fast in exploring the new version of transformers and how it works with GLiNER. I will work on adding support for the Transformers v5.

[Ingvarstep]

@elvismdev, the issue was that Transformers v5 initializes tokenizers differently from spm.model, and special tokens like bos and eos weren't added to the sequence. It was fixed in this PR: #326