dKhRr

oneinbox-qwen3.5-4b-lora

Deploy Dedicated

README

License: apache-2.0

Training Details

Base Model: Qwen/Qwen3.5-4B
Training Framework: Tinker Supervised LoRA Training
Training Set: 540 diverse synthetic emails curated to reflect realistic email structures, edge cases, urgent requests, and newsletters
Adapter Configuration: Rank $r = 32$ , Alpha $α = 32$ , targeting all-linear modules

Classification Labels & Schema

The model classifies an email into one of four labels:

must_act: Immediate action required (bills, account suspension notices, security alerts, calendar invitations with impending deadlines).
worth_a_look: High-value personal opportunities, relevant technical releases, developer meetups, or beta invites.
fyi: Informational notifications requiring no action (receipts, change password confirmations, closed issue updates).
noise: Marketing blasts, promotional discounts, routine automated newsletters, and social clutter.

Output JSON Format

json

{  "label": "<must_act | worth_a_look | fyi | noise>",  "why": "<under 10 words explanation>",  "deadline": "<YYYY-MM-DD or null>",  "summary": "<concise one-line summary>"}

Evaluation Benchmark

Evaluated on data/val.jsonl (60 held-out synthetic emails, 28 containing non-null gold deadlines). Latency was benchmarked in parallel with concurrency 10.

Table with columns: Variant, Label Accuracy, Macro-F1, JSON Validity, Deadline Exact-Match, p50 Latency, Cost / 1k Emails
Variant	Label Accuracy	Macro-F1	JSON Validity	Deadline Exact-Match	p50 Latency	Cost / 1k Emails
Base Qwen3.5-4B (0-shot)	52/60 (86.67%)	0.8645	60/60 (100.00%)	23/28 (82.14%)	2.58s	$0.1569
Base Qwen3.5-4B (3-shot)	56/60 (93.33%)	0.9328	60/60 (100.00%)	25/28 (89.29%)

Fine-Tuned Per-Label Breakdown

must_act: Precision 1.0000 | Recall 1.0000 | F1: 1.0000
worth_a_look: Precision 0.9333 | Recall 0.9333 | F1: 0.9333
fyi: Precision 1.0000 | Recall 0.9333 | F1: 0.9655
noise: Precision 0.9375 | Recall 1.0000 | F1: 0.9677

Honest Limitations

Synthetic Training Distribution: The 540 training emails and 60 validation emails are synthetically generated. While designed with realistic edge cases and varied phrasing, performance on real-world email distributions with noisy formatting or multipart HTML may vary.
Untested on Production Inboxes: The pipeline and fine-tuned model have not yet been evaluated on real-world corporate or personal inboxes due to privacy considerations.
Inference Pricing Assumption: The official Tinker pricing documentation lists base Qwen3.5-4B rates ( $0.33/ M p r o m ptt o k e n s,$ 1.005/M completion tokens) but does not provide a distinct rate for sampling fine-tuned checkpoints. The reported cost per 1,000 emails ($0.1412) assumes base model pricing.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

dKhRr

Model Tree

Base

Qwen/Qwen3.5-4B

Adapter

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality