ImAXISOG commited on
Commit
aab26f9
·
verified ·
1 Parent(s): d5c9d1d

Model card: exact Windows-file comparisons, training starting point, 8 GB not yet tested

Browse files
Files changed (1) hide show
  1. README.md +5 -4
README.md CHANGED
@@ -252,9 +252,9 @@ Keep every point of the draft and add nothing. Keep who does what, and keep ques
252
 
253
  ![How it was made: Gemma 4 E4B-it, writing lessons, LoRA training, merged and quantized, MLX and GGUF](https://huggingface.co/ImAXISOG/iris-gemma-4-e4b-it-mlx/resolve/main/assets/made.png)
254
 
255
- | **Starting point** | [google/gemma-4-E4B-it](https://huggingface.co/google/gemma-4-E4B-it), with thinking switched off in the chat template |
256
  |---|---|
257
- | Lessons | 10,200 distinct examples (13,439 rows: the newest Hindi lessons appear twice), 408 held back to check progress |
258
  | Where lessons came from | Drafts and answers written and reviewed with AI models; Hindi and Hinglish translations reviewed by AI reviewers who corrected tense, idiom and who-does-what. None is taken from the test drafts, and lessons that resembled a test draft were dropped. Not released. |
259
  | Method | LoRA on every layer: rank 16, scale 20, dropout 0.05; loss on the answers only (prompts masked) |
260
  | Schedule | One pass, 3,360 steps, batch 4, learning rate 5e-5 with cosine decay to 2e-6, 20 warm-up steps, up to 2,048 tokens |
@@ -268,14 +268,15 @@ Keep every point of the draft and add nothing. Keep who does what, and keep ques
268
  - **Drafts:** messages people type at work and to friends and family, written only for testing and never shown to
269
  whoever wrote the lessons: 50 English drafts (Fix and all five tones), 82 Hindi and Hinglish drafts for Fix, and
270
  240 tone answers on Hindi and Hinglish drafts.
271
- - **Answers:** temperature 0 with Iris's short prompts, through the same code the app uses.
 
272
  - **Judge:** an AI model, blind: the answers of all three models to a draft were shown together in shuffled order,
273
  with no model names, and judged against written rules.
274
  - **Fix "accepted":** no real mistake left and nothing changed beyond trivial tidying (a reworded phrase, an expanded
275
  short form or a lost casual voice counts against it). For Hindi drafts: an accurate, natural English version.
276
  - **Tones "would send":** the writer could send it in place of the draft: the same meaning, the tone asked for, and
277
  it reads naturally.
278
- - **The GGUF file** was judged on the answers that differ from the MLX model's. Fix and tones (English and Hindi) within about 2 points of it; Funny about 12 points lower (6 fewer good answers of 50).
279
 
280
  ## Limitations
281
 
 
252
 
253
  ![How it was made: Gemma 4 E4B-it, writing lessons, LoRA training, merged and quantized, MLX and GGUF](https://huggingface.co/ImAXISOG/iris-gemma-4-e4b-it-mlx/resolve/main/assets/made.png)
254
 
255
+ | **Starting point** | Google's [google/gemma-4-E4B-it](https://huggingface.co/google/gemma-4-E4B-it), as the 4-bit MLX copy [mlx-community/gemma-4-e4b-it-4bit](https://huggingface.co/mlx-community/gemma-4-e4b-it-4bit) (what Iris used before), with thinking switched off in the chat template |
256
  |---|---|
257
+ | Lessons | 10,200 distinct examples (13,439 rows: some lessons appear more than once), 408 held back to check progress |
258
  | Where lessons came from | Drafts and answers written and reviewed with AI models; Hindi and Hinglish translations reviewed by AI reviewers who corrected tense, idiom and who-does-what. None is taken from the test drafts, and lessons that resembled a test draft were dropped. Not released. |
259
  | Method | LoRA on every layer: rank 16, scale 20, dropout 0.05; loss on the answers only (prompts masked) |
260
  | Schedule | One pass, 3,360 steps, batch 4, learning rate 5e-5 with cosine decay to 2e-6, 20 warm-up steps, up to 2,048 tokens |
 
268
  - **Drafts:** messages people type at work and to friends and family, written only for testing and never shown to
269
  whoever wrote the lessons: 50 English drafts (Fix and all five tones), 82 Hindi and Hinglish drafts for Fix, and
270
  240 tone answers on Hindi and Hinglish drafts.
271
+ - **Answers:** through the same code the app uses. The trained models at temperature 0 with Iris's short prompts; the
272
+ untrained model as Iris ran it before, with Iris's longer prompts (worked examples included) and its usual settings.
273
  - **Judge:** an AI model, blind: the answers of all three models to a draft were shown together in shuffled order,
274
  with no model names, and judged against written rules.
275
  - **Fix "accepted":** no real mistake left and nothing changed beyond trivial tidying (a reworded phrase, an expanded
276
  short form or a lost casual voice counts against it). For Hindi drafts: an accurate, natural English version.
277
  - **Tones "would send":** the writer could send it in place of the draft: the same meaning, the tone asked for, and
278
  it reads naturally.
279
+ - **The GGUF file** was judged on the answers that differ from the MLX model's. Against the MLX model, counted over two judging rounds (MLX against the standard Q4_K_M, then that against this smaller file): English Fix 1 fewer good answer of 50, English tones 1 fewer of 200, Funny 6 fewer of 50, Hindi and Hinglish Fix 1 more of 82, Hindi and Hinglish tones 3 more of 240.
280
 
281
  ## Limitations
282