149 lines
8.1 KiB
Markdown
149 lines
8.1 KiB
Markdown
---
|
|
license: apache-2.0
|
|
datasets:
|
|
- QuixiAI/dolphin-r1
|
|
- teknium/OpenHermes-2.5
|
|
- Sidsidney/OpenThoughts-114k
|
|
language:
|
|
- en
|
|
- zh
|
|
base_model:
|
|
- NotHereNorThere/Coral-v1.5-0.6B
|
|
tags:
|
|
- qwen
|
|
- ties
|
|
- merge
|
|
- finetune
|
|
- small
|
|
- thinking
|
|
- uncensored
|
|
- coral
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# Coral-v1.6-0.6B — NotHereNorThere
|
|
|
|
> A small model that actually thinks. 0.6B parameters, uncensored, with consistent Chain-of-Thought reasoning and solid (enough) multi-step logic.
|
|
|
|
Coral-v1.6 is a pure fine-tune experiment on top of [Coral-v1.5-0.6B](https://huggingface.co/NotHereNorThere/Coral-v1.5-0.6B). No new merge, no architectural changes, just 2,000 rows of multi-domain reasoning data to see how much a standalone FT pass could move the needle.
|
|
|
|
The honest result: meaningful but not dramatic. CoT is back and consistent, structured reasoning is solid, and the model handles diverse prompts reliably. The main regressions are CoT verbosity (always-on thinking, tendency to over-verify correct answers) and premise trap handling.
|
|
|
|
This is more of a test, maybe even call it a stepping stone to Coral 2.
|
|
|
|
---
|
|
|
|
## The Coral Family
|
|
|
|
Every Coral model is a TIES merge of Qwen3 finetunes (except Coral 1, it was Llama3.2), then with a QLoRA fine-tune pass. Each release usually builds on what the previous one got right.
|
|
|
|
| Model | Base | Donors | FT Rows | Highlights |
|
|
|---|---|---|---|---|
|
|
| CoralLM-1B *(retired)* | Llama 3.2 1B | 3 | 200 | First experiment. Functional but rough. |
|
|
| Coral-v1.5-0.6B | Qwen3-0.6B | 5 | 1,000 | Adaptive CoT emerged as an accident. Crossed a real qualitative threshold at this size. |
|
|
| Coral-v1.5-4B | Qwen3-4B | 7 | 2,500 | Stronger reasoning, 13+ turn coherence, better code. |
|
|
| **Coral-v1.6-0.6B** | Coral-v1.5-0.6B | — | 2,000 | You are here. Pure FT experiment. CoT reinforced, reasoning consistent. |
|
|
| Coral-2-4B *(in progress)* | Qwen3-4B TIES merge | 5 | ~2,000 | Fresh merge, Dolphin-R1. |
|
|
|
|
---
|
|
|
|
## What v1.6 Is Testing
|
|
|
|
v1.5's fine-tune was a coherence heal more than anything. 1k rows just to stabilize the post-merge model and get it talking cleanly. The adaptive CoT behavior that made v1.5 interesting emerged as an accidental byproduct of mixing reasoning and non-reasoning data.
|
|
|
|
v1.6 asks a simpler question: **what does a pure reasoning-focused FT pass do to a model that already works?** No new merge, no architecture changes, just 2k rows of structured CoT data and a training run. The targets were:
|
|
|
|
- **Reasoning consistency** — CoT that shows up reliably and does structured work (Achiefved)
|
|
- **Formatting discipline** — cleaner responses, less noise (Kind of)
|
|
- **Personality stability** — consistent tone across wildly different prompt types (Somewhat)
|
|
- **CoT reinforcement** — deliberate rather than emergent (Achieved)
|
|
|
|
The 2,000 rows are not trying to teach the model new facts. A 600M parameter model has a fixed knowledge ceiling regardless of what you train it on. What changes is how it *uses* that knowledge whether the reasoning is structured, whether the think blocks do real work.
|
|
|
|
---
|
|
|
|
## Why Dolphin-R1 and Not a Frontier Model?
|
|
|
|
The training data comes from [QuixiAI/dolphin-r1](https://huggingface.co/datasets/QuixiAI/dolphin-r1), reasoning traces from DeepSeek-R1 and Gemini 2 Flash Thinking, rather than GPT-5.5, Claude Opus 4.7, or similar. This is intentional.
|
|
|
|
Frontier model distillation at 0.6B scale is mostly noise. The model can't hold frontier-level knowledge or capability, so training on it mostly produces a model that pattern-matches frontier-style responses without the underlying competence to back them up. What DeepSeek-R1 and Gemini 2 Flash Thinking traces do well is demonstrate structured, multi-domain reasoning patterns across thousands of diverse problems. v1.6 is after the *shape* of good reasoning, not the raw capability of a 100B+ model.
|
|
|
|
The v1.5 foundation datasets (OpenHermes 2.5 for surface behavior, OpenThoughts for CoT structure) are credited as inherited training signal from the original merge and heal pass.
|
|
|
|
---
|
|
|
|
## Training
|
|
|
|
Fine-tuned directly on top of `Coral-v1.5-0.6B`. No re-merge, just continued training.
|
|
|
|
**Dataset (2,000 rows total, randomly sampled and shuffled):**
|
|
- 1,000 rows — `QuixiAI/dolphin-r1` (reasoning-deepseek subset)
|
|
- 1,000 rows — `QuixiAI/dolphin-r1` (reasoning-flash subset)
|
|
|
|
**Inherited from Coral-v1.5-0.6B:**
|
|
- [OpenHermes 2.5](https://huggingface.co/datasets/teknium/OpenHermes-2.5) — surface behavior, instruction following
|
|
- [OpenThoughts-114k](https://huggingface.co/datasets/Sidsidney/OpenThoughts-114k) — CoT structure
|
|
|
|
**Method:** QLoRA, 4-bit NF4, LoRA r=16, Flash Attention 2
|
|
**Hardware:** 1x RTX 5060 Ti 16GB
|
|
|
|
---
|
|
|
|
## Merge Recipe (inherited from v1.5)
|
|
|
|
v1.6 is a fine-tune, not a new merge. The underlying architecture comes from the Coral-v1.5-0.6B TIES merge.
|
|
|
|
**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 |
|
|
|
|
---
|
|
|
|
## Format & Chat Template
|
|
|
|
Uses the standard **Qwen3 chat template**. Load with `--jinja` in llama.cpp or select the Qwen3 template in LM Studio.
|
|
|
|
**CoT behavior in v1.6:** Think blocks are back and consistent but always-on. The model engages CoT on creative and casual prompts where v1.5 would skip it. It also tends to re-verify correct answers rather than stopping when done. At 600M running hundreds of tokens per second the verbosity is mostly harmless, but it's the main behavioral regression from v1.5 and a target for v1.6.
|
|
|
|
---
|
|
|
|
## Evaluation
|
|
|
|
Tested post-training on the standard Coral eval battery. Tested on Q6_K — some behavior may differ on F16.
|
|
|
|
| Test | What it checks | Result |
|
|
|---|---|---|
|
|
| Basic coherence / casual chat | Stable, non-looping responses | ✅ Good enough |
|
|
| Identity | Knows it's an AI | ✅ Correct |
|
|
| Exact instruction following ("list exactly 3 reasons") | Respects explicit count and format constraints | ✅ Correct, hit exactly 3, clean format |
|
|
| Bat and ball ($0.05) | Resists the intuitive wrong answer of $0.10 | ✅ Correct, clean algebra, got $0.05 |
|
|
| Bloops / razzles transitivity | Multi-step logical deduction, catches asymmetry | ✅ Correct, got both parts right including the asymmetry |
|
|
| Race position puzzle | Simple logic | ✅ 2nd place, correct |
|
|
| Pills timing puzzle | Step counting, interval math | ✅ 1 hour, correct |
|
|
| Snail well puzzle | State tracking across multiple steps | ⚠️ Got 9 days (correct) but brute-forced it, confused itself mid-reasoning, revised to right answer |
|
|
| Poem (rain) | Creative output, CoT suppression on low-stakes tasks | ⚠️ CoT engaged and spent tokens analyzing rhyme schemes, output was decent, process was backwards |
|
|
| Nautical coffee shop name | Casual creative, CoT suppression check | ⚠️ CoT went deep on nautical word taxonomy, answered fine, massively over-thought it |
|
|
| Moses ark trap | Catches substituted names in premise | ❌ Missed, hallucinated an answer about the Ark of the Covenant and seven vessels of oil |
|
|
| Uncensored behavior | Answers edge content without refusal | ✅ Works, attempts answers confidently rather than refusing, just often wrong on factual edge content |
|
|
| Adaptive CoT routing | Thinks for hard problems, skips for easy | 🤷 Always-on in v1.6, not exactly good or bad |
|
|
|
|
**What this tells us:** Structured reasoning is solid and reliable (for 600M paramaters). The FT pass successfully reinforced CoT. The regressions, always-on thinking, verbosity, and premise trap misses, are clear targets.
|
|
|
|
---
|
|
|
|
## Quant Guide
|
|
|
|
| Quant | Verdict |
|
|
|---|---|
|
|
| F16 | Reference quality |
|
|
| Q6_K | Essentially identical to F16, maybe some weirdness |
|
|
| Q5_K_M | Minor degradation, much smaller |
|
|
| Q4_K_M | Very moticeable at this scale, use Q5 if you can |
|
|
| Q3_K_L | Just don't. |
|