PEFT
Safetensors
css
code-repair
lora
qwen2.5-coder
evidence-bound

Evidence-Bound CSS - Qwen2.5-Coder-1.5B LoRA v0.1

First controlled LoRA experiment over Evidence-Bound CSS v0.4.1.

Setup

  • Base model: Qwen/Qwen2.5-Coder-1.5B-Instruct
  • Train: 80 repair rows / 20 families
  • Held-out test: 20 rows / 5 unseen families
  • Train/test family overlap: 0
  • LoRA: r=16, alpha=32
  • Epochs: 3
  • Training GPU: NVIDIA L4
  • Primary evaluation: held-out browser success predicate

Results

  • Normalized exact CSS: base 4/20, tuned 0/20
  • Browser functional pass: base 8/20, tuned 4/20
  • Functional delta: -4 passes

Exact-string CSS match is intentionally secondary: semantically equivalent CSS can differ textually. The browser replay is the primary functional evidence.

See:

  • experiment_manifest.json
  • baseline_predictions.jsonl
  • tuned_predictions.jsonl
  • browser_eval_receipt.json

Status

This is an experiment, not a production model. A positive or negative result is preserved as evidence; no improvement is claimed unless the held-out browser metric improves.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ouroboroscollective/evidence-bound-css-qwen2.5-coder-1.5b-lora-v0.1

Adapter
(236)
this model

Dataset used to train ouroboroscollective/evidence-bound-css-qwen2.5-coder-1.5b-lora-v0.1