frostdev-ops/Rime-1.7B-GGUF

Rime 1.7B GGUF

Rime is the assistant in Glaze, the browser extension for Blackboard, Canvas and Moodle. Frostdev fine-tuned it from Liquid AI's LFM3-1.7B for Glaze's tasks, and Glaze runs it on the student's own computer in its WebGPU engine. Glaze downloads these files only when the student asks, and checks each one against its SHA-256.

File	Quant	Size	SHA-256
Rime-1.7B-Q8_0.gguf	Q8_0	1.85 GB	7520da22275730692ff54bf88dbd25309bc89fb8c9a7803dc7e18c02bda40999
Rime-1.7B-Q6_K.gguf	Q6_K	1.43 GB	1414d3501009228a4f332a49d91e0599da74275cca49b2ed98dd57a7bade3688
Rime-1.7B-Q4_K_M.gguf	Q4_K_M	1.10 GB	cdf235cba2228d58148a245846b3bad4324c626936f7981496a45d0b837ecea2

What it does

• Writing feedback against an assignment's rubric, with a grade band.
• Reviews of a draft (spelling, grammar, clarity), quoting the student's own words.
• Outlines and assignment requirements, citations and source lists from web pages, and reading a syllabus's grade weights.
• Study help that hints and checks a step without giving the answer away.
• Tool use over Glaze's Agents tools (to-do list, schedule), and short gray-text completions while the student writes.

It answers in English, Spanish, German, French, Dutch and Portuguese (Brazil).

Format

Glaze's prompts use LFM3's ChatML template with 16 added special tokens (ids 125033–125048): one per task, plus fences around page and document text so that text can't pass as instructions. The tuned context window is 4,096 tokens. Glaze runs every task with thinking off (an empty <think></think> block). Without Glaze's task tokens it behaves as a general ChatML chat model.

Evaluation

Glaze's own release gate compared Rime with stock LFM3-1.7B at the same quant, both run in Glaze's engine with Glaze's real task code. The gate requires Rime to beat stock on every metric at every quant, with JSON validity at 100% and no loss on grounding or prompt injection. Rime passed it at Q8_0, Q6_K and Q4_K_M. Selected metrics, Rime / stock:

Task	Metric	Q4_K_M	Q6_K	Q8_0
Gray-text completion	accepted	95% / 10%	95% / 39%	94% / 38%
Review	planted errors found	92% / 67%	93% / 62%	91% / 67%
Feedback	rubric criteria covered	99% / 54%	97% / 52%	100% / 22%
Outline	role accuracy	93% / 49%	93% / 54%	93% / 53%
Requirements	all four right	100% / 48%	100% / 61%	100% / 61%
Study help	step judged right	72% / 64%	74% / 64%	79% / 54%
Agent tools	task check passed	90% / 0%	95% / 0%	95% / 0%
Feedback	instructions planted in the draft obeyed	0% / 50%	0% / 50%	0% / 50%

The test sets are synthetic and held out from training. Quality scores came from an LLM judge against a rubric.

Training

LoRA fine-tuning (supervised, then preference optimization) on synthetic examples of Glaze's tasks. The synthetic data was generated, checked by Glaze's own task validators, and filtered for leaks from the test sets. Open datasets were also used, listed with their licenses in NOTICE.

Limitations

• It's a 1.7B model, so it gets things wrong. Feedback and grades are a second opinion, not a teacher's.
• Grading in Spanish, German and Dutch tends to under-score strong (A and B) drafts.
• Outside its six languages and Glaze's tasks, expect stock LFM3-1.7B quality or lower.

License

LFM Open License v1.0 (LICENSE), the base model's license. Commercial use by an organization with annual revenue of $10 million or more is not licensed. See NOTICE for what Frostdev changed and for the data sources' attributions.