PrevonFounder/prevon-pulse-mk6

Prevon Pulse MK6

Prevon Pulse is a 270M-parameter fine-tune of FunctionGemma that picks an app action and fills in its arguments for the Prevon notes app, entirely on device. It is trained to read the tool menu it is given (renamed, masked and shuffled tools), to answer NoMatchingAction when nothing fits, and to stop on its own after one call.

FunctionGemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms

Files

file	for
prevon-pulse-mk6.litertlm	LiteRT-LM: Android, iOS, Apple Silicon Mac, Windows
prevon-pulse-mk6-Q8_0.gguf	llama.cpp: Intel Mac
safetensors/	bf16 Hugging Face weights, tokenizer and config (source of the two files above)
training/	run settings, data hash and loss curve
SHA256SUMS.txt	checksums of every file

Training

Full fine-tune of google/functiongemma-270m-it on 196,597 rows covering 80 tools, 2 epochs, learning rate 3e-5, 32 rows per step, loss on the answer tokens only, trained on Tenstorrent Blackhole chips. The released weights are the mean of three independent runs (seeds 1, 2 and 3) that share the same starting model (a weight average: no extra training). See training/ for the settings, data hash and loss curves.

Evaluation

Scored on the exported .litertlm with the real LiteRT-LM engine (CPU). The test sets are model-written, were never used in training, and the raw result files are in eval/.

test	end to end	named right	wrongly refused	refusal ok	missed-tool declines	no answer
real3, one tool shown, native prompt (comparable with MK2 and MK3)	95.4%	99.8%	0.2%	93.0%	62.9%	0
real4: reminders, alarms, settings, suggestions, checklists (bucket menus)	91.2%	94.8%	2.8%	61.4%	0.0%	0
real7: the six newest tools and calendar questions (bucket menus)	88.8%	97.8%	0.0%	39.4%	0.0%	0
real8: the nine newest tools plus near-miss requests that must be declined (bucket menus)	89.2%	93.5%	1.8%	58.3%	0.0%	0
real8b: tools MK4 already knew, a regression check (bucket menus)	96.7%	96.7%	1.1%	0.0%	0.0%	0

Argument quality on real7 (the right values, not just the right tool):

mk6_soup3 real7 arguments
tool                           rows  named  exact  close  extra   (percent of NAMED-right rows)
AddEvent                         30  100.0   73.3   86.7    3.3
GetNoteStats                     29  100.0   96.6   96.6    0.0
GetWritingStats                  20  100.0   75.0   75.0    0.0
ListReminders                    12  100.0   91.7   91.7    0.0
ListTrash                        30   93.3  100.0  100.0    0.0
SetFolderIcon                    30   96.7   86.2   86.2    0.0
SetNoteIcon                      30   96.7   89.7   89.7    0.0
ALL                             181   97.8   87.6   89.8    0.6

per argument (close):
  AddEvent                     duration    75.0  of 4
  AddEvent                     time        96.7  of 30
  AddEvent                     what        86.7  of 30
  AddEvent                     where       60.0  of 5
  GetNoteStats                 id          96.6  of 29
  GetWritingStats              period      58.3  of 12
  ListReminders                time        87.5  of 8
  SetFolderIcon                icon        89.7  of 29
  SetFolderIcon                id          96.6  of 29
  SetNoteIcon                  icon        93.1  of 29
  SetNoteIcon                  id          96.6  of 29

first misses:
  AddEvent               | add an all day event for my birthday on march 3 | gold {"what": "birthday", "time": "march 3 all day"} | got {"duration": "all day", "time": "on march 3", "what": "birthday on march 3"}
  AddEvent   

Argument quality on real8:

mk6_soup3 real8 arguments
tool                           rows  named  exact  close  extra   (percent of NAMED-right rows)
ChangeEvent                      20  100.0   75.0   80.0    0.0
FileNotesToFolder                20   85.0   70.6   94.1    0.0
FindInFiles                      18  100.0   83.3   94.4    0.0
GetRecentChanges                 19  100.0   84.2   84.2    0.0
GetStatus                        17   76.5   92.3   92.3    0.0
ReadFileForAssistant             18  100.0   66.7   88.9    0.0
ReadLoginForAssistant            20  100.0   95.0  100.0    0.0
SetFolderColor                   21  100.0  100.0  100.0    0.0
ShowMe                           17   76.5   76.9  100.0    0.0
ALL                             170   93.5   83.0   92.5    0.0

per argument (close):
  ChangeEvent                  change      95.0  of 20
  ChangeEvent                  id          85.0  of 20
  FileNotesToFolder            folder     100.0  of 17
  FileNotesToFolder            which       94.1  of 17
  FindInFiles                  query       94.4  of 18
  GetRecentChanges             since       84.2  of 19
  GetStatus                    about       92.3  of 13
  ReadFileForAssistant         fileName    88.9  of 18
  ReadLoginForAssistant        title      100.0  of 20
  SetFolderColor               colorName  100.0  of 21
  SetFolderColor               id         100.0  of 21
  ShowMe                       what       100.0  of 13

first misses:
  GetStatus              | is my vault l

MK6 notes

MK6 is the mean of three independent runs (seeds 1, 2 and 3), trained on 196,597 rows over 80 tools, two epochs. It adds EditNoteLines (line-level edits to a note) and the tools MK5 did not know.

Official ship-rule gate (all real3 rows plus real4, CPU, real export): e2e 95.4, wrongly refused 0.2, refusal ok 93.0, miss declined 62.9 (MK5 soup: 95.1 / 0.5 / 93.0 / 68.2). It is weaker than MK5 at declining out-of-scope requests.

real9 (121 held-out EditNoteLines requests, bucket menu): 99.2 end to end. MK4 scored 9.1 and MK5 28.9 on the same rows (neither knew the tool).

Single-run MK6 models sometimes return no function call after export (up to 13 rows in one run); the three-run mean returned 1 across all nine gates. Scores on the older tools are within about 1 to 3 points of MK4; that gap was not tested against seed noise.

The sets are written and judged by language models, so real-world accuracy is not yet measured.

Use and limits

Built for the Prevon app's own prompt format; it is not a general chat model. Use is subject to the Gemma Terms of Use and Prohibited Use Policy (https://ai.google.dev/gemma/terms, https://ai.google.dev/gemma/prohibited_use_policy).

License

Gemma Terms of Use (ai.google.dev/gemma/terms) and Prohibited Use Policy (ai.google.dev/gemma/prohibited_use_policy) apply in full. Prevon adds a short set of terms: see PREVON_MODEL_TERMS.md. Short version: free for people, schools, researchers and non-profits; companies and anything that makes money need Prevon's written permission; competitors are never covered; rename it if you change it.