nixie-querygen-v2
A Mistral-7B-v0.1 fine-tuned on query generation task. Main use cases:
- synthetic query generation for downstream embedding fine-tuning tasks - when you have only documents and no queries/labels. Such task can be done with the nixietune toolkit, see the
nixietune.qgen.generate
recipe. - synthetic dataset expansion for further embedding training - when you DO have query-document pairs, but only a few. You can fine-tune the
nixie-querygen-v2
on existing pairs, and then expand your document corpus with synthetic queries (which are still based on your few real ones). Seenixietune.qgen.train
recipe.
The idea behind the approach is taken from the doqT5query model. See the original paper Rodrigo Nogueira and Jimmy Lin. From doc2query to docTTTTTquery.
Training data
We used 200k query-document pairs sampled randomly from a diverse set of IR datasets:
Flavours
This repo has multiple versions of the model:
- model-*.safetensors: Pytorch FP16 checkpoint, suitable for down-stream fine-tuning
- ggml-model-f16.gguf: GGUF F16 non-quantized llama-cpp checkpoint, for CPU inference
- ggml-model-q4.gguf: GGUF Q4_0 quantized llama-cpp checkpoint, for fast (and less precise) CPU inference.
Prompt formats
The model accepts the followinng prompt format:
<document next> [short|medium|long]? [question|regular]? query:
Some notes on format:
[short|medium|long]
and[question|regular]
fragments are optional and can be skipped.- the prompt suffix
query:
has no trailing space, be careful.
Inference example
With llama-cpp and Q4 model the inference can be done on a CPU:
$ ./main -m ~/models/nixie-querygen-v2/ggml-model-q4.gguf -p "git lfs track will \
begin tracking a new file or an existing file that is already checked in to your \
repository. When you run git lfs track and then commit that change, it will \
update the file, replacing it with the LFS pointer contents. short regular query:" -s 1
sampling:
repeat_last_n = 64, repeat_penalty = 1.100, frequency_penalty = 0.000, presence_penalty = 0.000
top_k = 40, tfs_z = 1.000, top_p = 0.950, min_p = 0.050, typical_p = 1.000, temp = 0.800
mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampling order:
CFG -> Penalties -> top_k -> tfs_z -> typical_p -> top_p -> min_p -> temp
generate: n_ctx = 512, n_batch = 512, n_predict = -1, n_keep = 0
git lfs track will begin tracking a new file or an existing file that is
already checked in to your repository. When you run git lfs track and then
commit that change, it will update the file, replacing it with the LFS
pointer contents. short regular query: git-lfs track [end of text]
Training config
The model is trained with the follwing nixietune config:
{
"train_dataset": "/home/shutty/data/nixiesearch-datasets/query-doc/data/train",
"eval_dataset": "/home/shutty/data/nixiesearch-datasets/query-doc/data/test",
"seq_len": 512,
"model_name_or_path": "mistralai/Mistral-7B-v0.1",
"output_dir": "mistral-qgen",
"num_train_epochs": 1,
"seed": 33,
"per_device_train_batch_size": 6,
"per_device_eval_batch_size": 2,
"bf16": true,
"logging_dir": "logs",
"gradient_checkpointing": true,
"gradient_accumulation_steps": 1,
"dataloader_num_workers": 14,
"eval_steps": 0.03,
"logging_steps": 0.03,
"evaluation_strategy": "steps",
"torch_compile": false,
"report_to": [],
"save_strategy": "epoch",
"streaming": false,
"do_eval": true,
"label_names": [
"labels"
]
}
License
Apache 2.0
- Downloads last month
- 68
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.