---
base_model: Qwen/Qwen3-1.7B
library_name: peft
language:
- en
tags:
- kanha
- qlora
- peft
- grounded-question-answering
---

# kanha.ai-1.7b-grounded-small-data-5ep-v1

This public repository contains the final QLoRA adapter checkpoint for the named Kanha small-data experiment. The customer-derived dataset is stored separately in the private `Kanha-AI/kanha-kanha.ai-1.7b-grounded-small-data-5ep-v1-dataset` repository.

## Provenance

- Base model: `Qwen/Qwen3-1.7B`
- Base model revision: `70d244cc86ccca08cf5af4e1e306ecf908b1ad5e`
- Tokenizer revision: `70d244cc86ccca08cf5af4e1e306ecf908b1ad5e`
- Dataset hash: `ced9b047c0d56cbe3bb96546db2903ff5e60e92ffdbad7d754d255422a810e4f`
- Epochs: 5.0
- NEFTune alpha: None

## Evaluation

The aggregate results were measured on the merged derivative of this adapter. Private prompts, answers, per-case identifiers, and outputs are excluded.

- Deterministic pass rate: 0.423077
- List recall: 0.671154
- Numbers recall: 0.964103
- Refusal rate: 0.038462
- Unsupported-value rate: 0.038462
- Cases: 26

## Inference contract

This adapter was trained and evaluated with retrieved source context and the native Qwen chat template with thinking disabled. Bare questions without source context are outside the evaluated contract.

## Immutable loading

The original training artifact is preserved exactly, so its `adapter_config.json` has a null revision. Do not resolve the base model from mutable HEAD. Load it explicitly:

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-1.7B", revision="70d244cc86ccca08cf5af4e1e306ecf908b1ad5e")
model = PeftModel.from_pretrained(base, "Kanha-AI/kanha-kanha.ai-1.7b-grounded-small-data-5ep-v1")
```