Refresh README.md

README.md

@@ -10,7 +10,6 @@ tags:
   - payments
   - ecommerce
   - lora
-  - qlora
   - function-calling
 language:
   - en
@@ -20,7 +19,7 @@ language:
 
 **Frontier-approaching e-commerce intent detection in a 1B model.** Given a live customer conversation, `zeroproof-ecommerce-1b` returns a single JSON object with the customer's payment intent and its structured details. It runs inline on live traffic, at a cost and latency where calling a frontier model is not an option.
 
-On a hard, held-out benchmark it reaches **77.5 macro-averaged intent-type accuracy, 96% of GPT-5's**, at **$0.18 per 1M output tokens** and about a second per response.
+On a hard, held-out benchmark it reaches **75.8 macro-averaged intent-type accuracy, 89% of GPT-5's**, at **$0.18 per 1M output tokens** and about a second per response.
 
 ## Model Details
 
@@ -30,7 +29,7 @@ On a hard, held-out benchmark it reaches **77.5 macro-averaged intent-type accur
 - **Model type:** E-commerce payment-intent classifier; structured JSON output over seven intent types
 - **Language:** English
 - **License:** Gemma (inherited from the base model)
-- **Finetuned from:** `google/gemma-3-1b-it`, via 4-bit QLoRA
+- **Finetuned from:** `google/gemma-3-1b-it`, via LoRA
 
 ### Model Sources
 
@@ -49,7 +48,7 @@ English-language e-commerce and payments only. It is a narrow intent classifier,
 
 ## Bias, Risks, and Limitations
 
-Trained and evaluated on synthetic role-play conversations from a multi-agent marketplace simulation, labeled in a separate model pass under a locked labeling policy; labels are model-assigned and carry residual noise, and live-traffic distribution may differ. Accuracy is currently strongest on refunds, exchanges, and transfers; checkout and bill-pay are the focus of the current data round.
+Trained and evaluated on synthetic role-play conversations from a multi-agent marketplace simulation, labeled in a separate model pass under a locked labeling policy; labels are model-assigned and carry residual noise, and live-traffic distribution may differ. Accuracy is currently strongest on refunds, subscriptions, and checkout; bill-pay is the focus of the current data round.
 
 ### Recommendations
 
@@ -90,7 +89,7 @@ The differentiator is the data. ZeroProof builds e-commerce intent training data
 
 #### Training Hyperparameters
 
-- **Method:** 4-bit QLoRA (via [unsloth](https://github.com/unslothai/unsloth))
+- **Method:** LoRA on the bf16 base (via [unsloth](https://github.com/unslothai/unsloth))
 - **Epochs:** 1
 - **Effective batch size:** 16
 - **Learning rate:** 2e-4
@@ -106,9 +105,9 @@ Held-out eval of 1,977 conversations, zero train/eval leakage, macro-averaged (e
 
 ![ZeroProof e-commerce intent evaluation: accuracy, per-action accuracy, and cost to serve](1b-v1.png)
 
-- Fine-tuning takes intent-type accuracy from **12 to 77.5** (macro), landing at **96% of GPT-5's 80.2**.
+- Fine-tuning takes intent-type accuracy from **12.1 to 75.8** (macro); on the frontier panel's own rows it scores **71.4, 89% of GPT-5's 80.2**.
 - On structured detail extraction it **exceeds the full frontier panel**: 64 against Sonnet 5's 56, Opus 4.8's 53, and GPT-5's 41.
-- On the no-action boundary it errs toward inaction: **9.2%** of true no-action conversations are marked actionable, with remaining errors concentrated in the safe direction (missed actions, not invented ones).
+- On the no-action boundary it errs toward inaction: **7.5%** of true no-action conversations are marked actionable, with remaining errors concentrated in the safe direction (missed actions, not invented ones).
 
 ## Technical Specifications