80 lines
2.3 KiB
Markdown
80 lines
2.3 KiB
Markdown
|
---
|
|||
|
|
license: apache-2.0
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
- zh
|
|||
|
|
base_model:
|
|||
|
|
- Qwen/Qwen3-0.6B
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
tags:
|
|||
|
|
- coral
|
|||
|
|
- base
|
|||
|
|
- merge
|
|||
|
|
- qwen
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Coral-v1.5-0.6B-raw
|
|||
|
|
|
|||
|
|
Honestly, just use [Coral-v1.5-0.6B](#) instead.
|
|||
|
|
|
|||
|
|
The raw TIES merge of 5 Qwen3-0.6B finetunes, **before any fine-tuning**. This is the unhealed version of [Coral-v1.5-0.6B](#), coherent enough to be interesting, weird enough to be honest about what a fresh merge looks like.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Behavior (pre-heal)
|
|||
|
|
|
|||
|
|
This model has not been fine-tuned post-merge. Expect:
|
|||
|
|
|
|||
|
|
- `<think>` blocks active but unfocused, reasoning style present, structure loose
|
|||
|
|
- Occasional identity hallucination (confidently invents its own origin)
|
|||
|
|
- Counting errors in lists (wrong numbering, unclosed formatting)
|
|||
|
|
- Context retention across turns is unreliable
|
|||
|
|
- Arithmetic attempts are structurally sound but digits can be wrong
|
|||
|
|
- Uncensored, refusals removed via abliterated donors
|
|||
|
|
|
|||
|
|
Not recommended for direct use.
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## 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
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Evaluation (pre-heal)
|
|||
|
|
|
|||
|
|
| Test | Result |
|
|||
|
|
|---|---|
|
|||
|
|
| Basic greeting | ✅ Clean response |
|
|||
|
|
| Identity | ❌ Hallucinated "Lysandra A." as creator |
|
|||
|
|
| Exact instruction following ("list 3 fruits") | ❌ Numbered 3/4/5 instead of 1/2/3 |
|
|||
|
|
| Context retention across turns | ❌ Re-greeted instead of recalling name |
|
|||
|
|
| Math (47 × 83) | ❌ Wrong answer, correct decomposition method |
|
|||
|
|
| Prime number function | ⚠️ Correct code, wrong example predictions |
|
|||
|
|
| One-sentence explanation | ✅ Stayed concise |
|
|||
|
|
| CoT routing | ⚠️ Active but unfocused, not yet adaptive |
|
|||
|
|
| Uncensored | ✅ Refusals removed |
|
|||
|
|
|
|||
|
|
For the healed version with all of the above fixed and quants, see [Coral-v1.5-0.6B](#).
|
|||
|
|
|
|||
|
|
---
|