prithivMLmods commited on
Commit
052b09d
·
verified ·
1 Parent(s): c279388

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +96 -1
README.md CHANGED
@@ -1,4 +1,99 @@
1
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2
  base_model:
3
  - domyn/Domyn-Edge-Teacher-Math
4
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ library_name: transformers
3
+ license: apache-2.0
4
+ pipeline_tag: text-generation
5
+ datasets:
6
+ - domyn/Domyn-Edge-Teacher-Math-Mix
7
+ language:
8
+ - en
9
+ - it
10
+ - es
11
+ - fr
12
+ - de
13
+ - pt
14
+ - ru
15
+ - zh
16
+ - ja
17
+ - ar
18
+ - nl
19
+ - id
20
+ - tr
21
  base_model:
22
  - domyn/Domyn-Edge-Teacher-Math
23
+ tags:
24
+ - text-generation-inference
25
+ - llama-cpp
26
+ - domyn
27
+ - domyn-edge
28
+ - teacher
29
+ - math
30
+ - grpo
31
+ - fully-open
32
+ ---
33
+
34
+ # **Domyn-Edge-Teacher-Math-GGUF**
35
+
36
+ > **Domyn Edge Teacher Math** (`domyn/Domyn-Edge-Teacher-Math`) is a fully open, 4B-parameter dense transformer built by Domyn on top of **Domyn Edge SFT** to serve as the specialized mathematics domain teacher in its Multi-Teacher On-Policy Distillation (MOPD) post-training pipeline. Released in bfloat16 as part of the second Domyn Edge bundle alongside its training dataset (`Domyn-Edge-Teacher-Math-Mix`), it was trained via asynchronous Group Relative Policy Optimization (GRPO) on math problems with verifiable final answers—using `math-verify` to award format-sensitive rewards (1.0 for a properly formatted correct answer, 0.5 for an unformatted correct answer, and 0 otherwise). Intended to score student rollouts token-by-token rather than operate as a general-purpose assistant, it boosts AIME 2026 avg@10 performance by +17.7 points over Domyn Edge SFT (from 61.3% to 79.0%) and outperforms Qwen3.5 4B on HMMT Feb 2025 (76.1%) and Feb 2026 (65.9%). Retaining the exact architecture, tokenizer, 13-language support, and `<think>` reasoning chat template of Domyn Edge SFT, it is released under an **Apache 2.0** license.
37
+
38
+ ## Model Files
39
+
40
+ | File Name | Quant Type | File Size | File Link | Description |
41
+ |-----------|------------|-----------|-----------|-------------|
42
+ | Domyn-Edge-Teacher-Math.BF16.gguf | BF16 | 7.87 GB | [Link](https://huggingface.co/prithivMLmods/Domyn-Edge-Teacher-Math-GGUF/blob/main/Domyn-Edge-Teacher-Math.BF16.gguf) | Full BF16 weights. Highest quality, largest file size. |
43
+ | Domyn-Edge-Teacher-Math.Q3_K_L.gguf | Q3_K_L | 2.16 GB | [Link](https://huggingface.co/prithivMLmods/Domyn-Edge-Teacher-Math-GGUF/blob/main/Domyn-Edge-Teacher-Math.Q3_K_L.gguf) | Lower quality but usable, good for low RAM availability. |
44
+ | Domyn-Edge-Teacher-Math.Q3_K_M.gguf | Q3_K_M | 2 GB | [Link](https://huggingface.co/prithivMLmods/Domyn-Edge-Teacher-Math-GGUF/blob/main/Domyn-Edge-Teacher-Math.Q3_K_M.gguf) | Low quality, smallest file size. |
45
+ | Domyn-Edge-Teacher-Math.Q4_K_M.gguf | Q4_K_M | 2.42 GB | [Link](https://huggingface.co/prithivMLmods/Domyn-Edge-Teacher-Math-GGUF/blob/main/Domyn-Edge-Teacher-Math.Q4_K_M.gguf) | Good quality, default size for most use cases, *recommended*. |
46
+ | Domyn-Edge-Teacher-Math.Q4_K_S.gguf | Q4_K_S | 2.31 GB | [Link](https://huggingface.co/prithivMLmods/Domyn-Edge-Teacher-Math-GGUF/blob/main/Domyn-Edge-Teacher-Math.Q4_K_S.gguf) | Slightly lower quality with more space savings, *recommended*. |
47
+ | Domyn-Edge-Teacher-Math.Q5_K_M.gguf | Q5_K_M | 2.81 GB | [Link](https://huggingface.co/prithivMLmods/Domyn-Edge-Teacher-Math-GGUF/blob/main/Domyn-Edge-Teacher-Math.Q5_K_M.gguf) | High quality, *recommended*. |
48
+ | Domyn-Edge-Teacher-Math.Q5_K_S.gguf | Q5_K_S | 2.75 GB | [Link](https://huggingface.co/prithivMLmods/Domyn-Edge-Teacher-Math-GGUF/blob/main/Domyn-Edge-Teacher-Math.Q5_K_S.gguf) | High quality with slight space savings, *recommended*. |
49
+ | Domyn-Edge-Teacher-Math.Q6_K.gguf | Q6_K | 3.23 GB | [Link](https://huggingface.co/prithivMLmods/Domyn-Edge-Teacher-Math-GGUF/blob/main/Domyn-Edge-Teacher-Math.Q6_K.gguf) | Very high quality, near perfect, *recommended*. |
50
+ | Domyn-Edge-Teacher-Math.Q8_0.gguf | Q8_0 | 4.18 GB | [Link](https://huggingface.co/prithivMLmods/Domyn-Edge-Teacher-Math-GGUF/blob/main/Domyn-Edge-Teacher-Math.Q8_0.gguf) | Extremely high quality, generally unneeded but max available standard quant. |
51
+
52
+ ## Q3_K_L Test Run
53
+
54
+ Tested `Domyn-Edge-Teacher-Math.Q3_K_L.gguf` on an NVIDIA RTX PRO 6000 Blackwell Server Edition GPU with `96 GB` of VRAM, using `llama-cli` with full GPU offloading.
55
+
56
+ ### Performance Metrics
57
+
58
+ | Metric | Value |
59
+ | --- | --- |
60
+ | Average Prompt Processing Speed | 1,109.73 tokens/s |
61
+ | Average Generation Speed | 267.61 tokens/s |
62
+ | Total Wall Time | 56.3 seconds |
63
+ | Context Length | 16,384 tokens |
64
+ | KV Cache (K) | `q8_0` |
65
+ | GPU | NVIDIA RTX PRO 6000 Blackwell Server Edition |
66
+ | GPU VRAM Capacity | 96 GB |
67
+ | Peak GPU VRAM Usage | 4.00 GiB |
68
+ | Peak GPU Utilization | 91% |
69
+ | GPU Layers Offloaded | 99 (`-ngl 99`) |
70
+ | Random Seed | 42 |
71
+ | Speculative Decoding | No |
72
+ | Test Runs | 5 |
73
+
74
+ ### Test Shots
75
+
76
+ | Shot | Test |
77
+ | --- | --- |
78
+ | Shot 1 | C: Insert Node at End of Singly Linked List |
79
+ | Shot 2 | Jogger and Train Passing Time |
80
+ | Shot 3 | Two Pipes Filling a Tank |
81
+ | Shot 4 | Students in Examination Rooms |
82
+ | Shot 5 | Father and Son Ages |
83
+
84
+ ### Run Inference
85
+
86
+ ```bash
87
+ ./build/bin/llama-cli \
88
+ -m /path/to/models/Domyn-Edge-Teacher-Math.Q3_K_L.gguf \
89
+ -ngl 99 \
90
+ -c 16384 \
91
+ --cache-type-k q8_0 \
92
+ --seed 42 \
93
+ -n 8192 \
94
+ -st \
95
+ -p "Your prompt here"
96
+ ```
97
+
98
+ *Performance results reflect this specific hardware and inference configuration. Actual throughput may vary depending on the prompt, model, software build, and runtime settings.*
99
+