---
library_name: peft
base_model: GritLM/GritLM-7B
base_model_relation: adapter
tags:
- lora
- recommendation
- agenttailor
---

# AgentTailor BookCrossing adapter

Author-supplied fine-tuned checkpoint for the BookCrossing setting of the
[AgentTailor project](https://github.com/Alex-bupt/AgentTailor).
The original checkpoint directory is named `ReFICR_qlora_bc`. This release
preserves the adapter weights byte-for-byte and replaces the training machine's
base-model path with the public `GritLM/GritLM-7B` identifier.

This is a LoRA adapter, requiring the separate GritLM-7B base model. It is not
a standalone merged model or the Stage-2 LLM agent. Use the AgentTailor code
for candidate preparation, agent scoring, evidence merging and evaluation.

## Checkpoint details

- Base model: [GritLM/GritLM-7B](https://huggingface.co/GritLM/GritLM-7B)
- LoRA rank: 64; alpha: 32; dropout: 0.05; bias: none
- Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
- 448 finite tensors, 167,772,160 adapter parameters
- Adapter file: 671,149,168 bytes
- SHA-256: `1dd900ae9752af6105ded55c67c1b53f8c67b3a4d97ad5bce113d7a86f8414e0`

The source folder is labeled QLoRA. The exported adapter itself is stored as
ordinary safetensors; this does not imply that the file contains a quantized
copy of the complete base model. The original auxiliary non-LoRA state dict
was inspected using `weights_only=True` and is empty, so it is omitted.
`training_base_config.json` records the supplied architecture for provenance;
load the base architecture and implementation from the base model repository.

## Load with GritLM and PEFT

Install the project's GritLM fine-tuning dependencies, including PyTorch,
Transformers, PEFT, Accelerate and GritLM, in a suitable GPU environment.
Use a local checkout of this adapter as `adapter_path`, or its Hugging Face
repository ID after publication:

```python
from gritlm import GritLM
from peft import PeftModel

adapter_path = './AgentTailor-BookCrossing'
model = GritLM('GritLM/GritLM-7B', mode='embedding', torch_dtype='auto')
model.model = PeftModel.from_pretrained(model.model, adapter_path).merge_and_unload()
model.model.eval()
embeddings = model.encode(['Example item description'], instruction='<|embed|>\n')
```

The adapter must actually be applied before encoding. Match the project's
profile/item formatting, instruction prefixes, pooling and attention settings
for experiment reproduction. Generic generation alone does not reproduce the
recommendation pipeline.

## Validation and scope

The release checks all adapter tensors for finite values, verifies the expected
32-layer LoRA coverage, loads the tokenizer locally, and checks copied weights
against SHA-256. Full GritLM-7B inference and benchmark reproduction have not
been run as part of this packaging step; no new performance claims are made.
Training logs, raw datasets and private machine paths are not distributed.

## Attribution

The code builds on GritLM and the supplied ReFICR fine-tuning implementation.
Consult the [base model card](https://huggingface.co/GritLM/GritLM-7B) for its
Apache-2.0 license and attribution. This card does not assign a new license to
the author-supplied adapter; no separate adapter license was supplied.