|
|||||||
![]() |
|
|
Thread Tools | Search this Thread |
|
|
#16 |
|
Leftutti
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() Posts: 649
Karma: 3671119
Join Date: Feb 2019
Location: Bavaria
Device: iPad Pro, Kobo Libra 2
|
With this gemma-3-1b I'm struggling getting the summary in German.
Edit: and there are strange duplicates Spoiler:
Last edited by Wiggo; 07-05-2026 at 05:58 AM. |
|
|
|
|
|
#17 | |
|
want to learn what I want
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() Posts: 2,347
Karma: 7990245
Join Date: Sep 2020
Device: Libra Colour
|
Quote:
I confirm this model can sometimes be sort of stubborn to produce output in languages other than English. And I don't recommend using local models at all - consumer hardware only supports highly quantized models and run at incomparably slower speed. Regarding the small duplicate content, what I see is: Chapter 2 "Prolog" summary: "Der junge Mann steht vor einem Fenster..." Highlight: "Sigille der Zukunft: Der junge Mann steht vor einem Fenster..." I'm taking a look, this could be due to an overly broad fallback used when model is outputting mixed-language or malformed responses that the parser mishandles. Thank you for posting the summary info! |
|
|
|
|
|
|
#18 | |||
|
Leftutti
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() Posts: 649
Karma: 3671119
Join Date: Feb 2019
Location: Bavaria
Device: iPad Pro, Kobo Libra 2
|
Quote:
Quote:
Quote:
|
|||
|
|
|
|
|
#19 |
|
want to learn what I want
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() Posts: 2,347
Karma: 7990245
Join Date: Sep 2020
Device: Libra Colour
|
I was trying to focus on another plugin I plan to release today but I found a fix for the language drift; new version will hopefully fully fix it, shortly.
|
|
|
|
|
|
#20 |
|
want to learn what I want
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() Posts: 2,347
Karma: 7990245
Join Date: Sep 2020
Device: Libra Colour
|
Version 1.3
- Fix language drift and redundancy on some highlights / key events section. - UI fixes Go for NVIDIA Free API. There are many good models such as Kimi 2.6, Llama instruct, Nemotrons Omni, Ultra and Super. The gpt oss is good too, but seems to have a more strict rate limit. |
|
|
|
|
|
#21 |
|
want to learn what I want
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() Posts: 2,347
Karma: 7990245
Join Date: Sep 2020
Device: Libra Colour
|
I've uploaded a hotfix for a minor issue. Latest version is 1.3.1.
|
|
|
|
|
|
#22 |
|
Leftutti
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() Posts: 649
Karma: 3671119
Join Date: Feb 2019
Location: Bavaria
Device: iPad Pro, Kobo Libra 2
|
Hi Comfy
CS 1.3.1. Spoiler:
Now it's a mix of German and English and the key event is sometimes not ignored, sometimes it is. Btw, I tried Nvidia but it took ages for the first chapter so I stopped it. |
|
|
|
|
|
#23 |
|
want to learn what I want
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() Posts: 2,347
Karma: 7990245
Join Date: Sep 2020
Device: Libra Colour
|
I'll take a look but this might take long this time. In the meantime, I suggest you try other models, preferrably not local. Try more Openrouter's free models like Laguna XS series and Liquid. I see there's a Hy3 there. The Tencent models are usually great and seem to be less rate-limit restricted.
https://openrouter.ai/models?q=free&order=newest |
|
|
|
|
|
#24 |
|
Member
![]() Posts: 24
Karma: 10
Join Date: Oct 2025
Device: Kindle
|
This is a very cool plugin! Thank you for sharing it. One question, can you provide an option to not include the run information when appending to comments?
I don't want to include this bit in the comments, especially for batch jobs: Code:
- Summary type: Flash overview - Source: E:\_shared\Non-Fiction\Iron Eyes Cody\Indian Talk Hand Signals of the Am (10454)\Indian Talk Hand Signals of the - Iron Eyes Cody.pdf - Format: pdf | Document type: nonfiction - Source language: English | Output language: same as source - Base paragraphs per chapter: 2 | Request pace used: 0s (free=1s, other=0s) - Reasoning strategy: auto | Chunk size: 1400 words - Provider: LMStudio | Requested model: google/gemma-4-e2b - Actual model: google/gemma-4-e2b - Requests: 2 (attempts: 2) - Duration: 0:28 | Generated: 2026-07-06 10:37 |
|
|
|
|
|
#25 | |
|
want to learn what I want
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() Posts: 2,347
Karma: 7990245
Join Date: Sep 2020
Device: Libra Colour
|
Quote:
|
|
|
|
|
|
|
#26 |
|
Leftutti
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() Posts: 649
Karma: 3671119
Join Date: Feb 2019
Location: Bavaria
Device: iPad Pro, Kobo Libra 2
|
OK, it’s the same as always in life – good work comes at a price.
I’ve just spent $10 and it’s quick and there’s no more rate limit. ChatGPT recommended me this one - Qwen3 Next 80B A3B Instruct |
|
|
|
|
|
#27 | |
|
want to learn what I want
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() Posts: 2,347
Karma: 7990245
Join Date: Sep 2020
Device: Libra Colour
|
Quote:
Yeah, they got me too.Qwens would often give me timeouts on Openrouter. |
|
|
|
|
|
|
#28 |
|
Leftutti
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() Posts: 649
Karma: 3671119
Join Date: Feb 2019
Location: Bavaria
Device: iPad Pro, Kobo Libra 2
|
![]() Oh, 34 chapters are only 0,11$ Edit: ChatGPT's additional prompt recommendation Spoiler:
Last edited by Wiggo; 07-06-2026 at 03:11 PM. |
|
|
|
|
|
#29 |
|
want to learn what I want
![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() ![]() Posts: 2,347
Karma: 7990245
Join Date: Sep 2020
Device: Libra Colour
|
Stick to the free models and relax privacy permissions, such as "Allow prompt usage" and the like. This expands model availability.
|
|
|
|
|
|
#30 |
|
Member
![]() Posts: 24
Karma: 10
Join Date: Oct 2025
Device: Kindle
|
I dislike paying for AI, it's so expensive with API calls. For API calls in the cloud, I use Cerebras.ai free tier due to it's simplictic API. It's always fast, but you have to pay attention to the rate limits for most free tiers.
For cerebras the limits are:
You can tell that most rate limits are descending in design so you have to actually calculate your rate based on the number of prompts you are sending and fit it into one of the rate limit categories. I'm working on comments for 1,000 books at approximately 2 prompts per books (observed testing), so ~2,000 prompts. The daily limit is 2400 requests or 1-request every 36-seconds (86400 sec/day / 2400 requests/day = 36). If I set a rate limiter with this number I can process all of my books within a 24-hour period. With this you could set a batch job with rate limiting, and just let it run on it's own until complete. This also keeps me under the 150-requests/hour and 5-requests/minute thresholds, so I know I won't be timed out by the platform during the run. You would need to request a rate limit feature for this plugin to do this though. For tasks like this I like to run a small local model. Most computers can run a small model in ollama. It's slow, but easy to setup, and you can let it pine away while you do other things. I've tested small Mistral variants and they work fine for this on plugin in ollama on an Intel i7 with 16-32GB of RAM. Small models are fine when you don't have a gaming video card or a system with unified RAM. |
|
|
|
![]() |
|
Similar Threads
|
||||
| Thread | Thread Starter | Forum | Replies | Last Post |
| [GUI Plugin] Streamer - RSVP Speed Reader plugin for Calibre | npapadacis | Plugins | 6 | 04-24-2026 03:00 AM |
| docxTranslat: plugin to import WinWord files into Sigil chapter by chapter | Chiavse | Plugins | 2 | 12-13-2023 10:38 AM |
| [GUI Plugin] Noosfere_util, a companion plugin to noosfere DB | lrpirlet | Plugins | 2 | 08-18-2022 03:15 PM |
| [GUI Plugin] Save Virtual Libraries To Column (GUI) | chaley | Plugins | 14 | 04-04-2021 05:25 AM |