amartomar683 commited on
Commit
b18482b
·
verified ·
1 Parent(s): 6e56e58

tiny-ncm weights (task6), tokenizer, tool registry, model card

Browse files
Files changed (4) hide show
  1. README.md +95 -0
  2. bpe.json +0 -0
  3. registry.json +24 -0
  4. tiny-ncm.pt +3 -0
README.md ADDED
@@ -0,0 +1,95 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-nc-4.0
3
+ language:
4
+ - en
5
+ library_name: pytorch
6
+ pipeline_tag: text-generation
7
+ tags:
8
+ - ncm
9
+ - coarse-to-fine
10
+ - vq-vae
11
+ - tool-use
12
+ - from-scratch
13
+ - tiny
14
+ datasets:
15
+ - HuggingFaceFW/fineweb-edu
16
+ - roneneldan/TinyStories
17
+ - wikimedia/wikipedia
18
+ - abisee/cnn_dailymail
19
+ - corbt/enron-emails
20
+ - Yale-LILY/aeslc
21
+ - HuggingFaceTB/everyday-conversations-llama3.1-2k
22
+ - HuggingFaceTB/smol-smoltalk
23
+ - Salesforce/xlam-function-calling-60k
24
+ ---
25
+
26
+ **Files:** `tiny-ncm.pt` (weights + config, 146 MB), `bpe.json` (text tokenizer), `registry.json` (tool list).
27
+ Run it with the code at https://github.com/amar683/tiny-ncm:
28
+
29
+ ```bash
30
+ git clone https://github.com/amar683/tiny-ncm && cd tiny-ncm
31
+ python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
32
+ hf download amartomar683/tiny-ncm tiny-ncm.pt --local-dir weights
33
+ .venv/bin/python chat.py --dry-run
34
+ ```
35
+
36
+ # Model card: tiny-ncm
37
+
38
+ ## Summary
39
+ - **Type:** NCM (coarse-to-fine, block-wise) language model for tool use, trained from scratch.
40
+ - **Size:** 36.4M parameters: generator 20.9M (d=384, 8 layers, 3 query heads / 1 KV head) and multiscale residual VQ
41
+ tokenizer 15.5M (5 levels, 1024-entry codebooks). Byte-level BPE vocabulary of 16,384. Context 4096.
42
+ - **Checkpoint:** `runs/task6` (released as `tiny-ncm.pt`, fp32, 146 MB). Code: https://github.com/amar683/tiny-ncm (MIT).
43
+ - **Hardware:** one MacBook (Apple M5, 24 GB), PyTorch MPS, bf16. About 12 hours of training in total (VQ tokenizer 1.3 h, generator 10.5 h).
44
+
45
+ ## Intended use
46
+ A personal assistant that turns everyday requests into tool calls (alarms, reminders, timers, calendar, read-only Gmail,
47
+ web and Wikipedia lookups, weather, notes, calculator) and makes small talk. Use it together with the app in
48
+ `assistant/`, which checks every tool call against the user's words. **Not** for coding, reasoning, medical, legal or
49
+ financial advice, or anything where a wrong answer matters.
50
+
51
+ ## Training data
52
+ | Data | License | Used for |
53
+ |---|---|---|
54
+ | [FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) (sample-10BT) | ODC-By 1.0 | general English |
55
+ | [TinyStories](https://huggingface.co/datasets/roneneldan/TinyStories) | CDLA-Sharing 1.0 | simple English |
56
+ | [Wikipedia](https://huggingface.co/datasets/wikimedia/wikipedia) (20231101.en) | CC BY-SA 3.0 / GFDL | text, lookup answers |
57
+ | [CNN/DailyMail](https://huggingface.co/datasets/abisee/cnn_dailymail) | Apache-2.0 | summaries |
58
+ | [Enron emails](https://huggingface.co/datasets/corbt/enron-emails), [AESLC](https://huggingface.co/datasets/Yale-LILY/aeslc) | not stated (public record) | email style, subject lines |
59
+ | [everyday-conversations](https://huggingface.co/datasets/HuggingFaceTB/everyday-conversations-llama3.1-2k), [smol-smoltalk](https://huggingface.co/datasets/HuggingFaceTB/smol-smoltalk) | Apache-2.0 | small talk (non-coding only) |
60
+ | [xLAM function-calling 60k](https://huggingface.co/datasets/Salesforce/xlam-function-calling-60k) | CC-BY-4.0, "research purposes only" | reading tool schemas |
61
+ | Synthetic tool conversations (`ncm/synth_tools.py`, `make_copy_data.py`) | made for this project | tool calls |
62
+
63
+ The weights were **not** trained on web pages collected by `harvest_web.py` (that experiment, `runs/web1`, was not adopted).
64
+
65
+ **Weights license: CC BY-NC 4.0.** Because of the xLAM terms and the data with no stated license, the weights are released for
66
+ **research and non-commercial use**, with credit to the datasets above. xLAM asks users to cite APIGen:
67
+ Liu et al., *APIGen: Automated Pipeline for Generating Verifiable and Diverse Function-Calling Datasets*, 2024.
68
+
69
+ ## Training
70
+ 1. BPE tokenizer (16k) on the mix.
71
+ 2. VQ tokenizer: 97.6% exact text reconstruction, 95% exact tool JSON.
72
+ 3. Generator pretraining at context 1024 (Muon + AdamW, warmup-stable-decay schedule).
73
+ 4. Chat/tool phases at 4096: synthetic tool data, then "copy-focused" data with fresh random names, times and places in every
74
+ example, then 2h45m of mostly English, then 2 × 1h of copy-focused data.
75
+ Details and every run: [docs/HOW_IT_WORKS.md](docs/HOW_IT_WORKS.md).
76
+
77
+ ## Evaluation
78
+ | Test | Model + app checks | Model alone |
79
+ |---|---|---|
80
+ | 500 new requests | 490 / 500 | n/a |
81
+ | 60 hard requests written before the fixes | 57 / 60 | n/a |
82
+ | 153 everyday requests, right tool | 153 / 153 | n/a |
83
+ | 153 everyday requests, every detail exact | 153 / 153 | 64 / 153 |
84
+ | Math (90) | 89 / 90 | n/a |
85
+ | Coding (40) | 5 / 40 | n/a |
86
+
87
+ ## Limitations and risks
88
+ - **Copying.** The model often garbles names, numbers and times ("London" → "Singapore"); the app fixes them from the
89
+ user's words. Without the app, expect about half of the tool calls to have a wrong detail.
90
+ - **English.** Free-form replies are rough; the app replaces replies that don't match the tool result.
91
+ - **Reasoning.** No multi-step math, no code.
92
+ - **Web answers** are picked from the top search results by word matching; they can be wrong or mix methods.
93
+ - **No safety training.** The model has no refusal or content filtering. The app asks before any action that sends
94
+ messages or calls someone, and never guesses those tools from a garbled name.
95
+ - **Privacy.** Everything runs locally. Web questions go to DuckDuckGo and the pages' sites; Gmail access is read-only.
bpe.json ADDED
The diff for this file is too large to render. See raw diff
 
registry.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": 1,
3
+ "tools": [
4
+ {"name": "set_alarm", "desc": "set an alarm", "args": {"time": "str", "label": "str?"}, "platforms": ["mac", "windows", "linux", "android", "ios"]},
5
+ {"name": "set_reminder", "desc": "create a reminder", "args": {"text": "str", "time": "str?"}, "platforms": ["mac", "windows", "linux", "android", "ios"]},
6
+ {"name": "set_timer", "desc": "start a countdown timer", "args": {"duration": "str", "label": "str?"}, "platforms": ["mac", "windows", "linux", "android", "ios"]},
7
+ {"name": "calendar_add", "desc": "add a calendar event", "args": {"title": "str", "time": "str", "duration": "str?"}, "platforms": ["mac", "android", "ios"]},
8
+ {"name": "calendar_list", "desc": "list calendar events", "args": {"day": "str"}, "platforms": ["mac", "android", "ios"]},
9
+ {"name": "list_emails", "desc": "list recent emails", "args": {"query": "str?", "n": "int?"}, "platforms": ["mac", "windows", "linux", "android", "ios"]},
10
+ {"name": "read_email", "desc": "read one email", "args": {"id": "str"}, "platforms": ["mac", "windows", "linux", "android", "ios"]},
11
+ {"name": "draft_email", "desc": "save an email draft", "args": {"to": "str", "subject": "str", "body": "str"}, "platforms": ["mac", "windows", "linux", "android", "ios"]},
12
+ {"name": "send_email", "desc": "send an email", "args": {"to": "str", "subject": "str", "body": "str"}, "platforms": ["mac", "windows", "linux", "android", "ios"], "confirm": true},
13
+ {"name": "lookup", "desc": "look up a topic in an encyclopedia", "args": {"topic": "str"}, "platforms": ["mac", "windows", "linux", "android", "ios"]},
14
+ {"name": "web_search", "desc": "search the web", "args": {"query": "str"}, "platforms": ["mac", "windows", "linux", "android", "ios"]},
15
+ {"name": "weather", "desc": "get the weather", "args": {"location": "str?", "when": "str?"}, "platforms": ["mac", "windows", "linux", "android", "ios"]},
16
+ {"name": "notes_add", "desc": "save a note", "args": {"text": "str"}, "platforms": ["mac", "windows", "linux", "android", "ios"]},
17
+ {"name": "notes_search", "desc": "search saved notes", "args": {"query": "str"}, "platforms": ["mac", "windows", "linux", "android", "ios"]},
18
+ {"name": "calculator", "desc": "evaluate arithmetic", "args": {"expression": "str"}, "platforms": ["mac", "windows", "linux", "android", "ios"]},
19
+ {"name": "get_time", "desc": "current time in a place", "args": {"place": "str"}, "platforms": ["mac", "windows", "linux", "android", "ios"]},
20
+ {"name": "open_app", "desc": "open an app", "args": {"app": "str"}, "platforms": ["mac", "windows", "android", "ios"]},
21
+ {"name": "call", "desc": "phone call a contact", "args": {"contact": "str"}, "platforms": ["android", "ios"], "confirm": true},
22
+ {"name": "sms", "desc": "send a text message", "args": {"to": "str", "message": "str"}, "platforms": ["android", "ios"], "confirm": true}
23
+ ]
24
+ }
tiny-ncm.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:085a893255dbe9e946f5099d8b4c6943cbd63f8e83b4c270e0b7b78b90fcaec7
3
+ size 145837082