Hugging Face's logo

blazeofchi
/
mempool-qwen-logits-orchestrator-smoke

Transformers
orchestration
routing
qwen
logits-head
Model card Files
xet
 Community
blazeofchi commited on
Commit
ed55de0
·
verified ·
1 Parent(s): 49fd2b1

Upload folder using huggingface_hub

Browse files
Files changed (9) hide show
  1. 20260628-qwen-small-logits-orchestrator-smoke-plan.json +40 -0
  2. 20260628-qwen-small-logits-orchestrator-split-smoke-plan.json +40 -0
  3. 20260628-qwen-training-readiness-py311.json +21 -0
  4. README.md +46 -0
  5. eval_report.json +82 -0
  6. heldout_eval_report.json +82 -0
  7. qwen_logits_heads.pt +3 -0
  8. train_eval_report.json +82 -0
  9. train_report.json +21 -0
20260628-qwen-small-logits-orchestrator-smoke-plan.json ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "can_train_here": true,
3
+ "config": {
4
+ "backend": "transformers",
5
+ "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
6
+ "batch_size": 1,
7
+ "epochs": 1,
8
+ "learning_rate": 0.0002,
9
+ "lora_rank": 0,
10
+ "max_length": 512,
11
+ "seed": 7,
12
+ "train_backbone": false
13
+ },
14
+ "dependency_status": {
15
+ "mlx": false,
16
+ "mlx_lm": false,
17
+ "torch": true,
18
+ "transformers": true
19
+ },
20
+ "record_count": 66,
21
+ "rows_output": "research/datasets/20260628-qwen-small-logits-orchestrator-smoke-rows.jsonl",
22
+ "schema_version": "mempool.qwen_logits_orchestrator_plan.v1",
23
+ "substrate": "/Users/paras/projects/mempool/research/datasets/20260628-m5-current-task-66task-substrate.jsonl",
24
+ "training_order": [
25
+ "freeze Qwen-small backbone",
26
+ "train worker/workflow/verifier/abstain heads on measured soft targets",
27
+ "compare held-out routing against the linear multi-head baseline",
28
+ "only enable LoRA/backbone updates after the heads beat the baseline"
29
+ ],
30
+ "worker_ids": [
31
+ "ollama-cloud-deepseek-v4-pro",
32
+ "ollama-cloud-glm-5.2",
33
+ "ollama-cloud-kimi-k2.7-code",
34
+ "ollama-cloud-qwen3-coder-480b"
35
+ ],
36
+ "workflow_labels": [
37
+ "direct",
38
+ "verify_then_fallback"
39
+ ]
40
+ }
20260628-qwen-small-logits-orchestrator-split-smoke-plan.json ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "can_train_here": true,
3
+ "config": {
4
+ "backend": "transformers",
5
+ "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
6
+ "batch_size": 1,
7
+ "epochs": 1,
8
+ "learning_rate": 0.0002,
9
+ "lora_rank": 0,
10
+ "max_length": 512,
11
+ "seed": 7,
12
+ "train_backbone": false
13
+ },
14
+ "dependency_status": {
15
+ "mlx": false,
16
+ "mlx_lm": false,
17
+ "torch": true,
18
+ "transformers": true
19
+ },
20
+ "prepared_rows": "research/datasets/20260628-qwen-small-logits-orchestrator-split-train.jsonl",
21
+ "record_count": 53,
22
+ "rows_output": "research/datasets/20260628-qwen-small-logits-orchestrator-split-train.jsonl",
23
+ "schema_version": "mempool.qwen_logits_orchestrator_plan.v1",
24
+ "training_order": [
25
+ "freeze Qwen-small backbone",
26
+ "train worker/workflow/verifier/abstain heads on measured soft targets",
27
+ "compare held-out routing against the linear multi-head baseline",
28
+ "only enable LoRA/backbone updates after the heads beat the baseline"
29
+ ],
30
+ "worker_ids": [
31
+ "ollama-cloud-deepseek-v4-pro",
32
+ "ollama-cloud-glm-5.2",
33
+ "ollama-cloud-kimi-k2.7-code",
34
+ "ollama-cloud-qwen3-coder-480b"
35
+ ],
36
+ "workflow_labels": [
37
+ "direct",
38
+ "verify_then_fallback"
39
+ ]
40
+ }
20260628-qwen-training-readiness-py311.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "transformers",
3
+ "backend_ready": true,
4
+ "dependency_status": {
5
+ "mlx": false,
6
+ "mlx_lm": false,
7
+ "torch": true,
8
+ "transformers": true
9
+ },
10
+ "machine": "arm64",
11
+ "platform": "macOS-26.3-arm64-arm-64bit",
12
+ "python_supported_for_torch": true,
13
+ "python_version": "3.11.14",
14
+ "ready_for_local_head_training": true,
15
+ "reasons": [],
16
+ "recommendations": [
17
+ "run a frozen-backbone head-training smoke before enabling LoRA or backbone updates"
18
+ ],
19
+ "require_gpu": false,
20
+ "schema_version": "mempool.qwen_training_readiness.v1"
21
+ }
README.md ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: Qwen/Qwen2.5-0.5B-Instruct
4
+ library_name: transformers
5
+ tags:
6
+ - orchestration
7
+ - routing
8
+ - qwen
9
+ - logits-head
10
+ pretty_name: mempool Qwen Logits Orchestrator Smoke
11
+ ---
12
+
13
+ # mempool Qwen Logits Orchestrator Split Smoke
14
+
15
+ This repository contains a split-smoke checkpoint for the `mempool`
16
+ Qwen-small logits-head orchestrator path with a deterministic held-out gate.
17
+
18
+ The checkpoint stores only the trained routing heads, not the Qwen base
19
+ model weights. Load the base model separately and attach the heads.
20
+
21
+ - Checkpoint: `qwen_logits_heads.pt`
22
+ - Training rows: `53`
23
+ - Final smoke loss: `4.2856650959770635`
24
+
25
+ Worker labels:
26
+
27
+ - `ollama-cloud-deepseek-v4-pro`
28
+ - `ollama-cloud-glm-5.2`
29
+ - `ollama-cloud-kimi-k2.7-code`
30
+ - `ollama-cloud-qwen3-coder-480b`
31
+
32
+ Train-row evaluation:
33
+
34
+ - Worker accuracy: `0.4716981132075472`
35
+ - Workflow accuracy: `0.5660377358490566`
36
+ - Mean worker loss: `1.4386708387788736`
37
+ - Mean workflow loss: `0.7223093554790501`
38
+
39
+ Held-out evaluation:
40
+
41
+ - Worker accuracy: `0.3076923076923077`
42
+ - Workflow accuracy: `0.7692307692307693`
43
+ - Mean worker loss: `1.7572260361451368`
44
+ - Mean workflow loss: `0.5600809453485104`
45
+
46
+ This is a smoke artifact, not a promoted production policy.
eval_report.json ADDED
@@ -0,0 +1,82 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "checkpoint": "/Users/paras/projects/mempool/research/models/20260628-qwen-small-logits-orchestrator-smoke/qwen_logits_heads.pt",
3
+ "mean_worker_loss": 1.4604794980788773,
4
+ "mean_workflow_loss": 0.41937757511605567,
5
+ "prediction_sample": [
6
+ {
7
+ "predicted_worker_id": "ollama-cloud-kimi-k2.7-code",
8
+ "predicted_workflow": "direct",
9
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
10
+ "target_workflow": "verify_then_fallback"
11
+ },
12
+ {
13
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
14
+ "predicted_workflow": "direct",
15
+ "target_worker_id": "ollama-cloud-kimi-k2.7-code",
16
+ "target_workflow": "direct"
17
+ },
18
+ {
19
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
20
+ "predicted_workflow": "direct",
21
+ "target_worker_id": "ollama-cloud-kimi-k2.7-code",
22
+ "target_workflow": "direct"
23
+ },
24
+ {
25
+ "predicted_worker_id": "ollama-cloud-deepseek-v4-pro",
26
+ "predicted_workflow": "direct",
27
+ "target_worker_id": "ollama-cloud-kimi-k2.7-code",
28
+ "target_workflow": "direct"
29
+ },
30
+ {
31
+ "predicted_worker_id": "ollama-cloud-glm-5.2",
32
+ "predicted_workflow": "direct",
33
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
34
+ "target_workflow": "verify_then_fallback"
35
+ },
36
+ {
37
+ "predicted_worker_id": "ollama-cloud-deepseek-v4-pro",
38
+ "predicted_workflow": "direct",
39
+ "target_worker_id": "ollama-cloud-deepseek-v4-pro",
40
+ "target_workflow": "direct"
41
+ },
42
+ {
43
+ "predicted_worker_id": "ollama-cloud-deepseek-v4-pro",
44
+ "predicted_workflow": "direct",
45
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
46
+ "target_workflow": "direct"
47
+ },
48
+ {
49
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
50
+ "predicted_workflow": "direct",
51
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
52
+ "target_workflow": "direct"
53
+ },
54
+ {
55
+ "predicted_worker_id": "ollama-cloud-kimi-k2.7-code",
56
+ "predicted_workflow": "direct",
57
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
58
+ "target_workflow": "direct"
59
+ },
60
+ {
61
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
62
+ "predicted_workflow": "direct",
63
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
64
+ "target_workflow": "verify_then_fallback"
65
+ }
66
+ ],
67
+ "record_count": 66,
68
+ "rows": "/Users/paras/projects/mempool/research/datasets/20260628-qwen-small-logits-orchestrator-smoke-rows.jsonl",
69
+ "schema_version": "mempool.qwen_logits_orchestrator_eval_report.v1",
70
+ "worker_accuracy": 0.3181818181818182,
71
+ "worker_ids": [
72
+ "ollama-cloud-deepseek-v4-pro",
73
+ "ollama-cloud-glm-5.2",
74
+ "ollama-cloud-kimi-k2.7-code",
75
+ "ollama-cloud-qwen3-coder-480b"
76
+ ],
77
+ "workflow_accuracy": 0.8636363636363636,
78
+ "workflow_labels": [
79
+ "direct",
80
+ "verify_then_fallback"
81
+ ]
82
+ }
heldout_eval_report.json ADDED
@@ -0,0 +1,82 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "checkpoint": "research/models/20260628-qwen-small-logits-orchestrator-split-smoke/qwen_logits_heads.pt",
3
+ "mean_worker_loss": 1.7572260361451368,
4
+ "mean_workflow_loss": 0.5600809453485104,
5
+ "prediction_sample": [
6
+ {
7
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
8
+ "predicted_workflow": "direct",
9
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
10
+ "target_workflow": "direct"
11
+ },
12
+ {
13
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
14
+ "predicted_workflow": "direct",
15
+ "target_worker_id": "ollama-cloud-kimi-k2.7-code",
16
+ "target_workflow": "direct"
17
+ },
18
+ {
19
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
20
+ "predicted_workflow": "verify_then_fallback",
21
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
22
+ "target_workflow": "verify_then_fallback"
23
+ },
24
+ {
25
+ "predicted_worker_id": "ollama-cloud-glm-5.2",
26
+ "predicted_workflow": "direct",
27
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
28
+ "target_workflow": "direct"
29
+ },
30
+ {
31
+ "predicted_worker_id": "ollama-cloud-kimi-k2.7-code",
32
+ "predicted_workflow": "direct",
33
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
34
+ "target_workflow": "direct"
35
+ },
36
+ {
37
+ "predicted_worker_id": "ollama-cloud-deepseek-v4-pro",
38
+ "predicted_workflow": "direct",
39
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
40
+ "target_workflow": "direct"
41
+ },
42
+ {
43
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
44
+ "predicted_workflow": "direct",
45
+ "target_worker_id": "ollama-cloud-kimi-k2.7-code",
46
+ "target_workflow": "direct"
47
+ },
48
+ {
49
+ "predicted_worker_id": "ollama-cloud-glm-5.2",
50
+ "predicted_workflow": "direct",
51
+ "target_worker_id": "ollama-cloud-glm-5.2",
52
+ "target_workflow": "direct"
53
+ },
54
+ {
55
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
56
+ "predicted_workflow": "verify_then_fallback",
57
+ "target_worker_id": "ollama-cloud-deepseek-v4-pro",
58
+ "target_workflow": "direct"
59
+ },
60
+ {
61
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
62
+ "predicted_workflow": "verify_then_fallback",
63
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
64
+ "target_workflow": "direct"
65
+ }
66
+ ],
67
+ "record_count": 13,
68
+ "rows": "research/datasets/20260628-qwen-small-logits-orchestrator-split-heldout.jsonl",
69
+ "schema_version": "mempool.qwen_logits_orchestrator_eval_report.v1",
70
+ "worker_accuracy": 0.3076923076923077,
71
+ "worker_ids": [
72
+ "ollama-cloud-deepseek-v4-pro",
73
+ "ollama-cloud-glm-5.2",
74
+ "ollama-cloud-kimi-k2.7-code",
75
+ "ollama-cloud-qwen3-coder-480b"
76
+ ],
77
+ "workflow_accuracy": 0.7692307692307693,
78
+ "workflow_labels": [
79
+ "direct",
80
+ "verify_then_fallback"
81
+ ]
82
+ }
qwen_logits_heads.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:670ec5385288ce86d7278a7df39469958fe495683acb9302d1f68acee70c164c
3
+ size 32921
train_eval_report.json ADDED
@@ -0,0 +1,82 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "checkpoint": "research/models/20260628-qwen-small-logits-orchestrator-split-smoke/qwen_logits_heads.pt",
3
+ "mean_worker_loss": 1.4386708387788736,
4
+ "mean_workflow_loss": 0.7223093554790501,
5
+ "prediction_sample": [
6
+ {
7
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
8
+ "predicted_workflow": "direct",
9
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
10
+ "target_workflow": "direct"
11
+ },
12
+ {
13
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
14
+ "predicted_workflow": "verify_then_fallback",
15
+ "target_worker_id": "ollama-cloud-kimi-k2.7-code",
16
+ "target_workflow": "direct"
17
+ },
18
+ {
19
+ "predicted_worker_id": "ollama-cloud-kimi-k2.7-code",
20
+ "predicted_workflow": "direct",
21
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
22
+ "target_workflow": "direct"
23
+ },
24
+ {
25
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
26
+ "predicted_workflow": "direct",
27
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
28
+ "target_workflow": "direct"
29
+ },
30
+ {
31
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
32
+ "predicted_workflow": "verify_then_fallback",
33
+ "target_worker_id": "ollama-cloud-kimi-k2.7-code",
34
+ "target_workflow": "direct"
35
+ },
36
+ {
37
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
38
+ "predicted_workflow": "verify_then_fallback",
39
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
40
+ "target_workflow": "direct"
41
+ },
42
+ {
43
+ "predicted_worker_id": "ollama-cloud-deepseek-v4-pro",
44
+ "predicted_workflow": "direct",
45
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
46
+ "target_workflow": "direct"
47
+ },
48
+ {
49
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
50
+ "predicted_workflow": "direct",
51
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
52
+ "target_workflow": "direct"
53
+ },
54
+ {
55
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
56
+ "predicted_workflow": "direct",
57
+ "target_worker_id": "ollama-cloud-qwen3-coder-480b",
58
+ "target_workflow": "direct"
59
+ },
60
+ {
61
+ "predicted_worker_id": "ollama-cloud-qwen3-coder-480b",
62
+ "predicted_workflow": "direct",
63
+ "target_worker_id": "ollama-cloud-deepseek-v4-pro",
64
+ "target_workflow": "direct"
65
+ }
66
+ ],
67
+ "record_count": 53,
68
+ "rows": "research/datasets/20260628-qwen-small-logits-orchestrator-split-train.jsonl",
69
+ "schema_version": "mempool.qwen_logits_orchestrator_eval_report.v1",
70
+ "worker_accuracy": 0.4716981132075472,
71
+ "worker_ids": [
72
+ "ollama-cloud-deepseek-v4-pro",
73
+ "ollama-cloud-glm-5.2",
74
+ "ollama-cloud-kimi-k2.7-code",
75
+ "ollama-cloud-qwen3-coder-480b"
76
+ ],
77
+ "workflow_accuracy": 0.5660377358490566,
78
+ "workflow_labels": [
79
+ "direct",
80
+ "verify_then_fallback"
81
+ ]
82
+ }
train_report.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "history": [
3
+ {
4
+ "epoch": 0,
5
+ "loss": 4.2856650959770635
6
+ }
7
+ ],
8
+ "output_dir": "research/models/20260628-qwen-small-logits-orchestrator-split-smoke",
9
+ "record_count": 53,
10
+ "schema_version": "mempool.qwen_logits_orchestrator_train_report.v1",
11
+ "worker_ids": [
12
+ "ollama-cloud-deepseek-v4-pro",
13
+ "ollama-cloud-glm-5.2",
14
+ "ollama-cloud-kimi-k2.7-code",
15
+ "ollama-cloud-qwen3-coder-480b"
16
+ ],
17
+ "workflow_labels": [
18
+ "direct",
19
+ "verify_then_fallback"
20
+ ]
21
+ }