File size: 3,263 Bytes
c6d82ea | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 | ---
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.
|