synthiumjp commited on
Commit
eef0eab
·
verified ·
1 Parent(s): e7ab07b

Upload folder using huggingface_hub

Browse files
README.md ADDED
@@ -0,0 +1,74 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model:
4
+ - Qwen/Qwen3-0.6B
5
+ - Qwen/Qwen3-1.7B
6
+ language: en
7
+ tags: [onnx, onnxruntime-genai, memory, notes, sourcedrecall]
8
+ ---
9
+
10
+ # sourcedrecall notes models (English, int4 ONNX)
11
+
12
+ Two small models for one job: given one conversation, write the lasting
13
+ facts the named person states about themselves, one short sentence per line,
14
+ or NONE. sourcedrecall runs them inside its own process on onnxruntime-genai,
15
+ on the CPU, with no server.
16
+
17
+ | file | base | download | per conversation (4 CPU threads) | peak memory |
18
+ |---|---|---|---|---|
19
+ | `sourcedrecall-notes-en-0.6b.tar.gz` (small) | Qwen3-0.6B | 0.51 GB | 2-3 s | 1.9 GB |
20
+ | `sourcedrecall-notes-en-1.7b.tar.gz` (standard) | Qwen3-1.7B | 1.3 GB | about 4.5 s | 3.3 GB |
21
+
22
+ They are optional. sourcedrecall stores and finds the user's own words
23
+ without them; with one, short notes are added, each kept only if its words
24
+ are in what the user wrote, and shown labelled as written by a model.
25
+
26
+ pip install 'sourcedrecall[notes]'
27
+ sourcedrecall-setup --notes # small
28
+ sourcedrecall-setup --notes standard # standard
29
+
30
+ ## How they were made
31
+
32
+ - Full fine-tune of the base model with mlx-lm, then int4 ONNX with the
33
+ onnxruntime-genai builder (k_quant_mixed).
34
+ - Teacher: Qwen3-14B wrote notes for 3,000 LongMemEval-S conversations (MIT)
35
+ and 1,500 SODA dialogues (CC BY 4.0), each SODA dialogue from both
36
+ speakers' sides; a note was kept only if most of its words are in the
37
+ person's own lines.
38
+ - Every training conversation got a random owner name and date format, and
39
+ its pronouns were made explicit (I/my -> the speaker, you/your -> the
40
+ listener), as sourcedrecall does before the model reads a conversation.
41
+ - Scripts: `tools/distil` in the sourcedrecall repository.
42
+
43
+ ## Measured (local Qwen3-14B reader and judge for every system)
44
+
45
+ LoCoMo, conversations 2-9 (1,307 questions):
46
+
47
+ | | overall | multi-hop | temporal | single-hop |
48
+ |---|---|---|---|---|
49
+ | sourcedrecall, no notes | 64.7% | 53.6% | 56.2% | 74.8% |
50
+ | with the small model's notes | 68.2% | 55.2% | 58.9% | 79.0% |
51
+ | with the standard model's notes | 70.3% | 59.0% | 57.4% | 81.6% |
52
+ | with Qwen3-14B's notes (a server) | 73.1% | 62.3% | 57.4% | 85.4% |
53
+ | Mem0 2.2.1 | 64.6% | 61.9% | 37.2% | 77.9% |
54
+
55
+ Notes on the LoCoMo development conversations, judged against their
56
+ conversation by Qwen3-14B: said by the person 80% (small), 87% (standard),
57
+ 93% (Qwen3-14B); gave the person the other speaker's facts 13%, 9%, 3%.
58
+ Held-out false-memory sets: v3 0 of 30 for both; v4, an old state given as
59
+ current, 3 of 44 (small) and 4 of 44 (standard), against 3 with Qwen3-14B's
60
+ notes and 2 without notes.
61
+
62
+ These answer scores come from local models and are not comparable with
63
+ published LoCoMo results, which use GPT-4-class models.
64
+
65
+ ## Limits
66
+
67
+ English only. They sometimes give the user something the other person said
68
+ ("has kids" from the other person's children). Neither closes
69
+ sourcedrecall's multi-hop gap to Mem0.
70
+
71
+ ## Attribution
72
+
73
+ SODA (Kim et al., 2023), CC BY 4.0. LongMemEval (Wu et al., 2024), MIT.
74
+ Qwen3 (Qwen team), Apache-2.0.
sourcedrecall-notes-en-0.6b.tar.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:db1b22779c3502665546780c91c5a8f3a1380f6e99570c05d3811e44b0c3d21e
3
+ size 511338454
sourcedrecall-notes-en-1.7b.tar.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0202bbb89468ee44fc397af37df1628dc55b7b46a414bb09b549101634ce5d91
3
+ size 1365037845