Mistral-7B-Instruct-v0.3 text-to-SQL, bf16
mistralai/Mistral-7B-Instruct-v0.3 fine-tuned on text-to-SQL, merged and left in bf16. It is the ceiling arm of a panel of 9 arms: every quantized arm below was made from this checkpoint and allocated the same byte budget, so their accuracies differ by method and not by size.
What this is
| base model | mistralai/Mistral-7B-Instruct-v0.3 |
| fine-tune | lora r=32, 2.0 epoch over 39,531 text-to-SQL conversations |
| the adapter it merged | VikramPal/mistral-7b-instruct-v0.3-lora |
| training data | gretelai/synthetic_text_to_sql, Salesforce/wikisql, b-mc2/sql-create-context |
| quantization | none -- this is the bf16 fine-tune every quantized arm was made from |
| size on disk | 13.500 GiB (16.0000 bits per parameter) |
| loads with | transformers |
Results
Execution match on 2,454 held-out text-to-SQL problems: the generated query is run against the schema and compared to the reference result set.
| arm | exec match | size | bits/param |
|---|---|---|---|
| bf16 | 78.16% | 13.500 GiB | 16.0000 |
| gptq_4b | 78.28% | 3.692 GiB | 4.3760 |
| awq_4b | 77.91% | 3.692 GiB | 4.3760 |
| dq_4b | 78.08% | 3.692 GiB | 4.3754 |
| gptq_3b | 6.68% | 2.858 GiB | 3.3869 |
| awq_3b | 74.16% | 2.858 GiB | 3.3869 |
| dq_3b | 75.22% | 2.857 GiB | 3.3859 |
| gptq_3b_asym_noao | 76.08% | 2.858 GiB | 3.3869 |
| gptq_3b_asym | 3.99% | 2.858 GiB | 3.3869 |
This arm, by evaluation source:
| eval source | exec match | items |
|---|---|---|
gretel | 77.02% | 818 |
spider | 63.57% | 818 |
wikisql | 93.89% | 818 |
How this arm compares
McNemar exact over the per-item hits, so every row is a paired test on the same problems in the same order. p (Holm) is step-down corrected within the family the panel declared, not within this card.
| comparison | delta (pts) | 95% CI | p | p (Holm) | verdict |
|---|---|---|---|---|---|
| 4b GPTQ vs bf16 | +0.12 | [-0.54, +0.79] | 0.81 | 1 | not separated |
| 4b AWQ vs bf16 | -0.24 | [-1.05, +0.56] | 0.621 | 1 | not separated |
| 4b DynQuant vs bf16 | -0.08 | [-0.83, +0.67] | 0.915 | 1 | not separated |
| 3b GPTQ vs bf16 | -71.48 | [-73.30, -69.65] | 0 | 0 | separated |
| 3b AWQ vs bf16 | -3.99 | [-5.29, -2.70] | 1.8e-09 | 9.01e-09 | separated |
| 3b DynQuant vs bf16 | -2.93 | [-4.16, -1.71] | 3.55e-06 | 1.42e-05 | separated |
What is not claimed
- The arms above are not all the same scheme. In this panel GPTQ runs symmetric with no activation reordering, asymmetric with no activation reordering, and asymmetric with group activation reordering; AWQ runs asymmetric with no activation reordering. DynQuant's quantizer is asymmetric and does not reorder columns, which is a property of the method rather than a recipe flag, so the panel records no scheme for its arms. Where a comparison above pairs a symmetric arm against an asymmetric one its delta spans two differences at once -- how the bits were allocated, and whether a zero point was stored per group -- so a large gap between those two arms is not on its own evidence about allocation. The comparison that isolates it is in this panel:
gptq_3bandgptq_3b_asym_noaoare the same method at the same byte anchor and differ in the scheme alone, so the difference between those two rows is the scheme and nothing else. - Storage, measured; throughput, not. The number reported here is bytes on disk and execution match. This card makes no claim about decode speed or peak VRAM against an fp16 baseline, because this panel did not measure either.
- One task. Execution match on held-out text-to-SQL is what was scored. It says nothing about how this arm behaves on anything else, and a quantization that holds one task can lose another.
Install
This checkpoint is plain bf16 and loads with transformers alone. It is the ceiling arm of a DynQuant panel: the quantized arms in the table above are this same fine-tune at a fraction of the size, and those need the package.
pip install dynquant
Source, format spec, and the allocator that produced their bit maps: https://github.com/kambojvikram/dynquant
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("VikramPal/mistral-7b-instruct-v0.3-bf16", device_map="cuda")
tokenizer = AutoTokenizer.from_pretrained("VikramPal/mistral-7b-instruct-v0.3-bf16")
Provenance
- panel model:
/workspace/runs/s4/mistral7b-v03.text2sql/merged - parameters counted: 7,248,023,552
- byte target: --
- fine-tune: 2472 steps, train loss 0.0540, 3.8 h
- fine-tune commit:
5959fe04b7db0512eaf7567d77a43f3df5860651 - evaluation: 2,454 problems in 10 min