---
library_name: transformers
model_name: corallm-1b-healed
tags:
- generated_from_trainer
- sft
- trl
- llama
- ties
- corallm
- openhermes
- uncensored
- 1b
licence: license
license: apache-2.0
datasets:
- teknium/OpenHermes-2.5
language:
- en
base_model:
- meta-llama/Llama-3.2-1B-Instruct
- bunnycore/Llama-3.2-1B-General-Best
- ai-nexuz/llama-3.2-1b-instruct-fine-tuned
- EpistemeAI/Reasoning-Llama-3.2-1B-Instruct-v1.2
pipeline_tag: text-generation
---
# CoralLM-1B

CoralLM-1B after a short instruction-tuning pass to clean up the scattered
behavior of the raw merge — tighter answers, cleaner endings, less rambling.

## Merge

- **Method:** TIES (`density 0.5` per model, `normalize: true`)
- **Base / anchor:** `meta-llama/Llama-3.2-1B-Instruct`
- **Components:**
  | Model | Weight | Brings |
  |---|---|---|
  | `EpistemeAI/Reasoning-Llama-3.2-1B-Instruct-v1.2` | 0.30 | reasoning, CoT |
  | `ai-nexuz/llama-3.2-1b-instruct-fine-tuned` | 0.30 | math, coding, logic |
  | `bunnycore/Llama-3.2-1B-General-Best` | 0.25 | general , creative, uncensored |

## Heal pass (training)

- **Method:** QLoRA (4-bit NF4, LoRA r=16)
- **Data:** 200 rows sampled from `teknium/OpenHermes-2.5`
- **Epochs:** 3 · **max_len:** 1024 · **grad_accum:** 8 · packing on
- **Loss:** stable ~1.5 throughout (learning without overfitting the 200 rows)

## Tests in Files

## Assessment

The heal pass did its job: output is tighter, the poem lost its essay-style
afterword, and answers end when they should. Reasoning ability is unchanged,
just more confidently expressed, which cuts both ways (cleaner right answers,
more assertive wrong ones). The 200-row pass tuned behavior without touching the
merged capabilities underneath.

**Safety:** the heal pass did not restore refusal behavior. Still effectively
uncensored, will attempt unsafe requests.