TR-3 study models: main compressed models and matched-budget baselines

Layout: / / . Method families are tr3 (our 3-site tensor ring + sparse residual), tt3 (3-site tensor train ablation), svd_llm, svd_llm_v2 and asvd_llm (matched-budget baselines). Stages are heal (KD-distilled, deployable) and unheal (raw post-compression, for measuring the healing delta).

method stage model params (B) licence KD result (probe)
tr3 heal qwen2.5-7b 5.187 qwen-research (non-commercial) see dir
tt3 heal llama2-7b 4.021 llama2 see dir
tt3 heal qwen2.5-3b 2.388 qwen-research (non-commercial) see dir
tt3 unheal llama2-7b 4.021 llama2
tt3 unheal qwen2.5-3b 2.388 qwen-research (non-commercial)
tr3 heal llama2-7b 4.023 llama2 MMLU 43.21*
tr3 heal llama2-13b 7.740 llama2 see dir
tr3 heal qwen2.5-3b 2.078 qwen-research (non-commercial) see dir
tr3 unheal llama2-13b 7.740 llama2
tr3 unheal llama2-7b (off-cluster archive) llama2
svd_llm heal llama2-7b 4.022 llama2 MMLU 39.47
svd_llm heal llama2-13b 7.738 llama2 MMLU 46.54
svd_llm unheal llama2-7b 4.022 llama2
svd_llm unheal llama2-13b 7.738 llama2
svd_llm_v2 ‡ heal llama2-7b 4.023 llama2 MMLU 40.22
svd_llm_v2 ‡ heal llama2-13b 7.740 llama2 MMLU 44.96
svd_llm_v2 ‡ unheal llama2-7b 4.023 llama2
svd_llm_v2 ‡ unheal llama2-13b 7.740 llama2
asvd_llm heal llama2-7b 4.023 llama2 MMLU 35.09
asvd_llm heal llama2-13b 7.736 llama2 MMLU 42.46
asvd_llm unheal llama2-7b 4.023 llama2
asvd_llm unheal llama2-13b 7.736 llama2

* full-benchmark lm-eval 0.4.12 result; SVD probe columns are limit-40 healing probes here — each dir's README/provenance carries the precise protocol.

‡ Official SVD-LLM V2 code was never released; svd_llm_v2 artifacts are our faithful implementation of the published algorithms (per-layer ranks document the allocation).

⚠️ These are NOT standard transformers checkpoints

tr3/** stores ring cores (load with each dir's tensor_formats.py); svd_llm*/** stores low-rank factor pairs (load with each dir's lowrank_linear.py). Weight files are uniformly named state_dict_fp16.pt. from_pretrained() will not work on this repo.

Benchmark results

All numbers below were produced by us under a single locked protocol: lm-eval-harness 0.4.12, fp16, seed 1234; MMLU 5-shot, HellaSwag 10-shot, everything else 0-shot; full benchmarks, no subsampling. Every baseline was re-run at the same parameter budget and the same 60M-token KD recipe as our model rather than quoted from its paper, so the rows within a block are directly comparable.

Perplexity is token-level over fixed non-overlapping 2048-token windows, following SVD-LLM's evaluater.py. It is not lm-eval's word-level perplexity; the two differ by the exponent N_tok/N_word (1.4156 for LLaMA-2 on WikiText-2) and must never be placed in the same column.

-- marks a cell not yet measured; it is a placeholder, not a zero or a failure.

Perplexity (lower is better)

Model Method Params (B) WikiText-2 C4
LLaMA-2-7B-chat dense FP16 6.738 6.94 9.48
TR-3 (ours) 4.023 8.34 11.10
SVD-LLM 4.022 11.49 15.09
SVD-LLM V2 ‡ 4.023 11.03 14.73
ASVD 4.023 11.22 14.18
TT-3 (open chain) 4.021 12.52 15.09
LLaMA-2-13B-chat dense FP16 13.016 6.11 8.52
TR-3 (ours) 7.740 6.98 9.65
SVD-LLM 7.738 10.15 13.46
SVD-LLM V2 ‡ 7.740 9.72 13.09
ASVD 7.736 9.13 11.88
LLaMA-2-7B (base) dense FP16 6.738 5.47 7.29
TR-3 (ours) 4.023 7.28 9.58
SVD-LLM 4.022 10.17 --
SVD-LLM V2 ‡ 4.023 9.73 --
ASVD 4.023 10.04 12.48
LLaMA-2-13B (base) dense FP16 13.016 4.88 6.74
TR-3 (ours) 7.740 6.18 8.51
ASVD 7.746 7.97 10.44
SVD-LLM 7.738 9.14 --
SVD-LLM V2 ‡ 7.740 8.74 --
Qwen2.5-3B dense FP16 3.086 8.56 --
TR-3 (ours) 2.078 11.75 --
TT-3 (open chain) 2.077 -- --
Matrix low-rank † 2.077 -- --
Qwen2.5-7B dense FP16 7.616 7.46 --
TR-3 (ours) 5.187 9.51 --

Zero-shot accuracy (%)

Model Method BoolQ ARC-C ARC-E LAMBADA PIQA Winogrande MathQA Avg
LLaMA-2-7B-chat dense FP16 80.58 43.26 68.39 69.82 77.04 68.35 28.48 62.27
TR-3 (ours) 77.09 38.48 65.11 66.52 75.52 66.30 27.50 59.50
SVD-LLM 67.98 32.76 57.20 56.76 71.55 62.19 25.49 53.42
SVD-LLM V2 ‡ 68.90 34.30 59.30 57.29 71.65 61.25 25.66 54.05
ASVD 69.69 33.28 57.03 55.15 71.33 60.14 24.56 53.03
TT-3 (open chain) 63.67 32.68 52.06 55.77 71.22 58.80 25.23 51.35
LLaMA-2-13B-chat dense FP16 81.96 49.91 73.40 72.79 79.33 72.85 32.86 66.16
TR-3 (ours) 80.52 43.86 70.71 70.74 76.82 70.01 29.92 63.22
SVD-LLM 73.64 35.41 61.03 60.59 72.74 64.48 25.93 56.26
SVD-LLM V2 ‡ 75.14 36.35 61.99 62.47 73.34 66.14 27.14 57.51
ASVD 74.89 36.95 64.60 61.60 74.05 65.19 26.37 57.66
LLaMA-2-7B (base) dense FP16 79.33 45.05 73.86 73.65 78.73 69.46 28.41 64.07
TR-3 (ours) 70.95 39.16 65.24 68.02 75.35 64.48 25.66 58.41
SVD-LLM 68.62 33.19 58.71 55.58 71.16 61.33 24.19 53.25
SVD-LLM V2 ‡ 68.53 33.96 60.06 56.49 72.31 62.83 25.59 54.25
ASVD 59.54 32.17 57.83 57.83 71.49 60.14 23.92 51.85
LLaMA-2-13B (base) dense FP16 82.11 48.81 76.47 76.54 80.25 72.45 31.46 66.87
TR-3 (ours) 76.21 44.11 72.43 71.74 78.51 68.67 29.65 63.05
ASVD 74.19 38.74 68.14 66.10 75.73 65.90 25.36 59.17
SVD-LLM 70.34 36.01 64.94 60.94 73.78 65.82 24.69 56.64
SVD-LLM V2 ‡ 71.13 38.14 67.30 61.67 73.83 65.67 25.66 57.63
Qwen2.5-3B dense FP16 80.09 48.04 72.90 66.12 78.29 69.38 35.61 64.35
TR-3 (ours) 76.39 43.17 69.36 60.51 75.24 64.09 34.17 60.42
TT-3 (open chain) 66.88 31.14 56.19 49.16 70.40 59.12 27.64 51.50
Matrix low-rank † 71.35 36.43 59.55 50.51 72.14 61.88 28.41 54.33
Qwen2.5-7B dense FP16 86.42 54.86 80.93 69.77 80.36 70.56 40.37 69.04
TR-3 (ours) 84.68 51.11 73.82 64.84 77.64 70.24 42.01 66.33

Few-shot / generative (%)

Model Method MMLU (5s) HellaSwag (10s) TriviaQA (5s) GSM8K (5s)
LLaMA-2-7B-chat dense FP16 47.22 78.63 57.18 22.59
TR-3 (ours) 43.21 70.75 36.93 9.40
SVD-LLM 39.47 57.65 16.18 2.05
SVD-LLM V2 ‡ 40.26 58.82 18.62 0.68
ASVD 34.65 59.46 18.02 2.27
TT-3 (open chain) 34.25 56.38 14.48 0.76
LLaMA-2-13B-chat dense FP16 53.55 82.07 65.20 35.25
TR-3 (ours) 49.69 76.46 47.81 24.11
SVD-LLM 44.18 63.41 21.57 0.15
SVD-LLM V2 ‡ 44.77 64.99 24.52 0.15
ASVD 41.43 68.01 29.71 7.51
LLaMA-2-7B (base) dense FP16 45.97 78.95 64.39 13.12
TR-3 (ours) 39.49 71.11 39.75 6.52
SVD-LLM 32.74 58.36 16.43 1.59
SVD-LLM V2 ‡ 34.33 59.36 19.34 1.97
ASVD 26.58 61.21 20.69 1.29
LLaMA-2-13B (base) dense FP16 55.54 82.14 70.69 23.28
TR-3 (ours) 50.46 77.04 50.96 14.33
ASVD 42.44 70.15 35.50 5.23
SVD-LLM 42.70 63.80 21.87 4.25
SVD-LLM V2 ‡ 44.67 65.00 24.49 4.93
Qwen2.5-3B dense FP16 66.42 75.26 45.78 10.31
TR-3 (ours) 56.02 66.39 27.61 41.32
TT-3 (open chain) 41.97 51.65 11.86 9.78
Matrix low-rank † 45.44 57.16 11.51 15.85
Qwen2.5-7B dense FP16 74.27 81.37 53.17 75.74
TR-3 (ours) 65.43 75.20 40.79 66.34

Matrix low-rank on Qwen2.5-3B is our own internal matrix-factorisation ablation, not a published method.

‡ Official SVD-LLM V2 code was never released; these are our faithful implementation of the published algorithms. Its close agreement with the official SVD-LLM v1 rows (within ~1.5 pp on every zero-shot task) is the evidence that the re-implementation is sound.

GSM8K on Qwen2.5 uses lm-eval's strict-match filter, which is fragile to output formatting on this model family. Under strict-match the dense Qwen2.5-3B teacher scores 10.31 while its compressed student scores 41.32; under flexible-extract the same teacher scores 63.46. The Qwen GSM8K column should therefore not be read as a capability ranking. LLaMA-2 rows are unaffected (strict and flexible agree to <0.2 pp on the dense teachers).

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support