Open-weight model
svd-safety-llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80
by Park Jeesup/svd-safety-llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80
meta-llama/Llama-2-7b-chat-hf compressed with Basis Sharing using the authors' code then recovered by LoRA on the coefficients alone.
Runs On
What it takes to serve svd-safety-llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80 (6.7B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 13.5 GB | 16.2 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 6.7 GB | 8.1 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 3.4 GB | 4.0 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Sep 18, 2026.
Model Card
meta-llama/Llama-2-7b-chat-hf compressed with Basis Sharing using the authors' code then recovered by LoRA on the coefficients alone. (the shared-basis fit) -> LoRA on the coefficients, bases frozen -> merge C' = C + (alpha/r)BA -> fold W = C' @ B to dense 2048 tokens, seed 42 (upstream hard-codes seed 2023; this project calibrates every method at one seed) 2 epochs, lr 0.0001, batch 64, on yahma/alpaca-cleaned Only the coefficients were trained. The shared and per-layer bases are frozen and bit-identical to the compressed model, so every weight keeps rank <= k, each group still shares one basis, and the parameter budget survives recovery exactly. This is not Basis Sharing's own LoRA…
Excerpt from the card by Park, licensed llama2.
Configuration
- Architecture
- LlamaForCausalLM
- Context length (tokens)
- 4,096
- Layers
- 32
- Hidden size
- 4,096
- Feed-forward size
- 11,008
- Attention heads
- 32
- Key/value heads
- 32
- Head dimension
- 128
- Vocabulary size
- 32,000
- RoPE base
- 10000
- Stored precision
- bfloat16
- Model type
- llama
Identity and Version
- Repository
- Jeesup/svd-safety-llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80
- Publisher
- Park
- Task
- Not stated by the source
- Modality
- Other
- Library
- Not stated by the source
- Parameters
- 6.7B parameters
- Languages
- low-rank
- Revision
- 0b941bc5322061f13729b6896f5f195c52bc1950
- First published
- 2026-09-18
- Last updated
- 2026-09-18
Files and Weights
39 files, 13.5 GB in total. The weights are 3 files totalling 13.5 GB in safetensors.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model-00001-of-00003.safetensors | Weights | 4.9 GB | dd805b77e4bd |
| model-00002-of-00003.safetensors | Weights | 4.9 GB | 702ee0f66496 |
| model-00003-of-00003.safetensors | Weights | 3.6 GB | 680fe1dc27e4 |
| build_manifest.json | Configuration | 2.7 KB | — |
| compression.json | Configuration | 16.1 KB | — |
| config.json | Configuration | 765 B | — |
| generation_config.json | Configuration | 188 B | — |
| hf_upload.json | Configuration | 358 B | — |
| model.safetensors.index.json | Configuration | 23.9 KB | — |
| safety/overrefusal.json | Configuration | 8.5 KB | — |
| safety/refusal.json | Configuration | 614.2 KB | — |
| safety/strongreject_summary.json | Configuration | 685 B | — |
| safety/summary.json | Configuration | 658 B | — |
| special_tokens_map.json | Configuration | 414 B | — |
| utility/ppl.json | Configuration | 190 B | — |
| utility/zeroshot/__lustre__jeesup0103__svd-safety__results_finetuned__llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80/results_2026-09-18T09-52-21.133360.json | Configuration | 24.4 KB | — |
| README.md | Documentation | 3.5 KB | — |
| safety/advbench_direct.jsonl | Other | 462.6 KB | — |
| safety/advbench_pool.jsonl | Other | 562.8 KB | — |
| safety/harmbench_judgments.jsonl | Other | 79.7 KB | — |
| safety/overrefusal_records.jsonl | Other | 2.7 MB | — |
| safety/per_prompt.csv | Other | 485.3 KB | — |
| safety/per_prompt.jsonl | Other | 567.1 KB | — |
| safety/strongreject_direct.jsonl | Other | 371.8 KB | — |
| safety/strongreject_harmbench_judgments.jsonl | Other | 49.1 KB | — |
| safety/strongreject_per_prompt.csv | Other | 386.5 KB | — |
| safety/strongreject_per_prompt.jsonl | Other | 435.8 KB | — |
| safety/strongreject_pool.jsonl | Other | 434.5 KB | — |
| utility/zeroshot/__lustre__jeesup0103__svd-safety__results_finetuned__llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80/samples_arc_challenge_2026-09-18T09-52-21.133360.jsonl | Other | 2.2 MB | — |
| utility/zeroshot/__lustre__jeesup0103__svd-safety__results_finetuned__llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80/samples_arc_easy_2026-09-18T09-52-21.133360.jsonl | Other | 4.1 MB | — |
| utility/zeroshot/__lustre__jeesup0103__svd-safety__results_finetuned__llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80/samples_hellaswag_2026-09-18T09-52-21.133360.jsonl | Other | 42.6 MB | 9b6cb52b52dc |
| utility/zeroshot/__lustre__jeesup0103__svd-safety__results_finetuned__llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80/samples_mathqa_2026-09-18T09-52-21.133360.jsonl | Other | 7.5 MB | — |
| utility/zeroshot/__lustre__jeesup0103__svd-safety__results_finetuned__llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80/samples_openbookqa_2026-09-18T09-52-21.133360.jsonl | Other | 647.5 KB | — |
| utility/zeroshot/__lustre__jeesup0103__svd-safety__results_finetuned__llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80/samples_piqa_2026-09-18T09-52-21.133360.jsonl | Other | 2.2 MB | — |
| utility/zeroshot/__lustre__jeesup0103__svd-safety__results_finetuned__llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80/samples_winogrande_2026-09-18T09-52-21.133360.jsonl | Other | 1.2 MB | — |
| .gitattributes | Repository | 1.7 KB | — |
| tokenizer.json | Tokenizer | 1.8 MB | — |
| tokenizer.model | Tokenizer | 499.7 KB | 9e556afd4421 |
| tokenizer_config.json | Tokenizer | 1.6 KB | — |
License and Download
- License
- llama2
- Access
- Open weights, no gate
- Download size
- 13.5 GB
Released by Park through its official repository on Hugging Face.
Built From
- Derived from meta-llama/Llama-2-7b-chat-hf
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 13.5 GB |
| 16-bit | 13.5 GB |
| 8-bit | 6.7 GB |
| 4-bit | 3.4 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About svd-safety-llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80
How much GPU memory does svd-safety-llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80 need?
About 16.2 GB at 16-bit and 4 GB at 4-bit: the weights (6.7B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run svd-safety-llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80 on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
What license is svd-safety-llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80 released under?
llama2, as its publisher declares it. Read the license text before commercial use.
What is svd-safety-llama2_7b_chat_up_basis_coeff_finetuned_keep_0p80's context length?
4,096 tokens, from the maximum position embeddings in its published configuration.