102 lines
3.3 KiB
Markdown
102 lines
3.3 KiB
Markdown
---
|
||
license: apache-2.0
|
||
datasets:
|
||
- Sidsidney/OpenThoughts-114k
|
||
- teknium/OpenHermes-2.5
|
||
language:
|
||
- en
|
||
- zh
|
||
base_model:
|
||
- Qwen/Qwen3-0.6B
|
||
tags:
|
||
- qwen
|
||
- ties
|
||
- merge
|
||
- finetune
|
||
- small
|
||
- thinking
|
||
- uncensored
|
||
- coral
|
||
pipeline_tag: text-generation
|
||
---
|
||
# Coral-v1.5-0.6B by NotHereNorThere
|
||
|
||
A 0.6B parameter uncensored generalist with **adaptive Chain-of-Thought reasoning**, it decides on its own whether a question needs thinking or not. Built from a 5-donor TIES merge of Qwen3-0.6B finetunes, healed with a 1k row fine-tune pass.
|
||
|
||
Part of the **Coral-v1.5** model family, which adds to the original CoralLM series (Llama 3.2 1B based). Coral-v1.5 moves to Qwen3 architecture for native `<think>` support and significantly improved base capability.
|
||
|
||
---
|
||
|
||
## What makes it interesting
|
||
|
||
- **Adaptive CoT at 0.6B** — the model routes dynamically: simple questions get instant answers, complex reasoning tasks trigger `<think>` blocks. This accidently emerged from the fine-tune data mix rather than being explicitly trained.
|
||
- **Uncensored** — refusal behavior has been removed via two abliterated donors. It just answers things.
|
||
- **Correct arithmetic** — passes basic math with clean step-by-step working.
|
||
|
||
---
|
||
|
||
## Merge Recipe
|
||
|
||
**Method:** TIES
|
||
**Base:** `Qwen/Qwen3-0.6B`
|
||
**Tool:** [mergekit](https://github.com/arcee-ai/mergekit)
|
||
|
||
| Donor | Role | Weight | Density |
|
||
|---|---|---|---|
|
||
| `reaperdoesntknow/Qwen3-0.6B-Distilled-30B-A3B-Thinking-SFT` | Thinking / reasoning | 0.30 | 0.5 |
|
||
| `MihaiPopa-1/Qwen-3-0.6B-Claude-4.7-Opus-Distilled` | Claude-style CoT | 0.30 | 0.5 |
|
||
| `suayptalha/Qwen3-0.6B-Code-Expert` | Code | 0.25 | 0.5 |
|
||
| `DavidAU/Qwen3-0.6B-heretic-abliterated-uncensored` | De-alignment | 0.15 | 0.5 |
|
||
| `huihui-ai/Huihui-Qwen3-0.6B-abliterated-v2` | De-alignment | 0.15 | 0.5 |
|
||
|
||
```yaml
|
||
base_model: Qwen/Qwen3-0.6B
|
||
merge_method: ties
|
||
dtype: bfloat16
|
||
parameters:
|
||
normalize: true
|
||
int8_mask: true
|
||
```
|
||
|
||
---
|
||
|
||
## Fine-tune
|
||
|
||
Post-merge heal pass to fix coherence, identity, counting, and context retention. Also reinforces when to use CoT vs when to answer directly.
|
||
|
||
- **500 rows** — [OpenHermes 2.5](https://huggingface.co/datasets/teknium/OpenHermes-2.5) (simple QA + instruction following)
|
||
- **500 rows** — [OpenThoughts](https://huggingface.co/datasets/open-thoughts/OpenThoughts-114k) (reasoning with CoT)
|
||
- **Method:** QLoRA + Flash Attention 2
|
||
- **Total:** 1,000 rows, randomly sampled and shuffled
|
||
|
||
The 50/50 split between non-CoT and CoT data is seemingly what produced the adaptive routing behavior.
|
||
|
||
---
|
||
|
||
## Evaluation
|
||
|
||
Tested post-heal on the following:
|
||
|
||
| Test | Result |
|
||
|---|---|
|
||
| Basic greeting | ✅ Clean, friendly, no loops |
|
||
| Identity | ✅ Identifies as AI assistant |
|
||
| Exact instruction following ("list 3 fruits") | ✅ Correct count and formatting |
|
||
| Context retention across turns | ✅ Recalled user name correctly |
|
||
| Math (47 × 83) | ✅ Correct (3,901) with clean working |
|
||
| Prime number function | ✅ Correct implementation and examples |
|
||
| One-sentence explanation | ✅ Stayed concise, no yapping |
|
||
| Adaptive CoT routing | ✅ Emergent, skips think for simple, uses think for complex |
|
||
| Uncensored | ✅ Refusals removed |
|
||
|
||
|
||
---
|
||
|
||
## Quant Guide
|
||
| Quant | Quality |
|
||
|---|---|
|
||
| F16 | Star of the show, best |
|
||
| Q6 | Should match F16 |
|
||
| Q5 | Starts degrading |
|
||
| Q4 | What could you run this on that's that bad |
|
||
| Q3 | Don't | |