Modelpublic

ermiaazarkhalili/FastContext-4B-RL_base-SFT-Fable5-GGUF

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes50downloads

Model Card

FastContext-4B-RL_base-SFT-Fable5-GGUF

GGUF quantizations of a LoRA fine-tune of microsoft/FastContext-1.0-4B-RL (no longer available on the Hub), supervised fine-tuned on ermiaazarkhalili/Fable-5-Complete-2M-Clean (private).

Quantized from `ermiaazarkhalili/FastContext-4B-RL_base-SFT-Fable5`. See that repository for the full-precision weights.


Base model	`microsoft/FastContext-1.0-4B-RL` (no longer available on the Hub)
Training data	`ermiaazarkhalili/Fable-5-Complete-2M-Clean` (private)
Method	LoRA supervised fine-tuning via Unsloth + TRL

Available quantizations

File	Size
`fastcontext-4b-rl_base-sft-fable5.q4_k_m.gguf`	2.50 GB
`fastcontext-4b-rl_base-sft-fable5.q5_k_m.gguf`	2.89 GB
`fastcontext-4b-rl_base-sft-fable5.q8_0.gguf`	4.28 GB

Usage

llama.cpp

bash

huggingface-cli download ermiaazarkhalili/FastContext-4B-RL_base-SFT-Fable5-GGUF fastcontext-4b-rl_base-sft-fable5.q4_k_m.gguf --local-dir .
llama-cli -m fastcontext-4b-rl_base-sft-fable5.q4_k_m.gguf -p "Explain gradient checkpointing in two sentences." -n 256

Ollama

bash

echo 'FROM ./fastcontext-4b-rl_base-sft-fable5.q4_k_m.gguf' > Modelfile
ollama create fastcontext-4b-rl_base-sft-fable5-gguf -f Modelfile
ollama run fastcontext-4b-rl_base-sft-fable5-gguf

Training configuration

Setting	Value
LoRA rank (r)	16
LoRA alpha	16
Learning rate	0.0002
Epochs	2
Effective batch size	8 (2 x 4 grad accum)
Max sequence length	4096
Base precision	4-bit (QLoRA)
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`

Observed training loss

Measured from our SLURM logs for this configuration. These are training-loss observations only — no downstream benchmark evaluation has been run on this model, so they should not be read as a quality claim.

SLURM job	Steps	First loss	Final loss
`53225525`	94,254	1.1631	0.8687

Limitations

—No benchmark evaluation has been run on this checkpoint. The only reported numbers are training-loss observations.
—Inherits the biases, knowledge cutoff and failure modes of the base model.
—Fine-tuned on a single instruction-following dataset; behaviour outside that distribution is untested.
—LoRA adapters were merged into the base weights, so the merged model cannot be detached from this fine-tune.

Reproducing

Trained by notebooks/fable_distillation_fastcontext-4b-rl_fable_unsloth.ipynb, executed non-interactively with papermill on a SLURM H100 partition (Unsloth + TRL, LoRA).

Card generated from the training run's own configuration and logs by `scripts/generate_hub_model_card.py`.