---
license: apache-2.0
language:
  - en
base_model: Qwen/Qwen2.5-1.5B-Instruct
base_model_relation: adapter
library_name: peft
pipeline_tag: text-generation
tags:
  - qlora
  - peft
  - governed-agent
  - retrieval
  - brain-navigator
  - grounded-only
  - proposal-only
  - research-only
  - szl-holdings
  - khipu
  - abstain-retrain
szl:
  doctrine: v11-LOCKED
  lean: "749/14/163"
  lambda: "Conjecture 1 — advisory, never a theorem"
  artifact_class: ADAPTER
  publication_eligible: true
  autonomy_eligible: false
  original_signed_weights: SZLHOLDINGS/SZL-Khipu-1.5B
---

# SZL-Khipu-1.5B-abstain

QLoRA **adapter** retrain of the existing Khipu line to raise in-memory abstain
oversample (ABSTAIN_OVERSAMPLE=4 → 32 abstain vs 15 navigate). Proposal-only.
Λ = Conjecture 1. Doctrine v11 LOCKED 749/14/163.

| | |
|---|---|
| **Base (canonical)** | [`Qwen/Qwen2.5-1.5B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) |
| **Runtime train** | `unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit` (same Qwen2.5-1.5B weights, 4-bit) |
| **Relation** | `adapter` (PEFT / Unsloth QLoRA) |
| **License** | Apache-2.0 |
| **Does NOT overwrite** | [`SZLHOLDINGS/SZL-Khipu-1.5B`](https://huggingface.co/SZLHOLDINGS/SZL-Khipu-1.5B) signed weights |
| **This is NOT** | the Chaski Qwen3.5 lock |

## Evaluation

**Status: MEASURED this job** (in-process port of `eval_khipu.py`, temperature 0, held-out never in gradients).

| split | k/n |
|---|---|
| plan-valid | 11 / 11 |
| grounding (eval.jsonl navigate) | 5 / 5 |
| abstain (adversarial.jsonl) | 3 / 6 |
| hallucinated citations | 0 |

Prior published original (`SZLHOLDINGS/SZL-Khipu-1.5B`) MEASURED abstain was **2/6** (blocker). This repo does not overwrite those signed weights. Counts above are this run only. Do not derive a leaderboard score from k/n on n=11.

## Training

- Unsloth QLoRA, seed 11, lr 0.0002, adamw_8bit, `train_on_responses_only`, Trackio
- LoRA r=32 α=64, epochs=45, ga=2, batch=1, constant_with_warmup
- ABSTAIN_OVERSAMPLE=4 (in-memory only; committed files unchanged)
- Train files: `train.jsonl` (15 navigate) + `train.abstain.jsonl` (8 rows × 4)
- Held-out: `eval.jsonl` (5) + `adversarial.jsonl` (6) — never in gradients
- finalTrainLoss (REPORTED string): `0.0172`
- adapter sha256 (safetensors bytes this job): `e44d53f29f2d443598e06d6c0441557fd3a5010888c7aa97b56ec3c0e050d349`

## Intended use

Supply a query + candidate Brain node **handles**. The adapter proposes a JSON
plan (`NAVIGATE` or `ABSTAIN`) per `khipu.schema.json`. A controller outside
the weights validates and resolves content. **Proposal-only. Not autonomous.**

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

base_id = "Qwen/Qwen2.5-1.5B-Instruct"
tok = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, "SZLHOLDINGS/SZL-Khipu-1.5B-abstain")
```

## Limitations

- Synthetic routing-policy harness, not live-Brain navigation skill.
- Small denominators (5 navigate / 6 abstain held-out).
- Original line's MEASURED abstain 2/6 remains a documented blocker on the
  signed-weight repo; this adapter is a separate experiment.