aaie-gft-full-fg-8k

AAIE GFT Full FG 8K

moinsaj/aaie-gft-full-fg-8k is a 354M-parameter AAIE research prototype for structured educational feedback generation. It is the standalone, full-weight epoch-2 checkpoint from two epochs of supervised feedback-generation training. It is not a LoRA adapter: download this repository alone to load the model.

Scope

The model was trained to generate this response shape from a complete AAIE assignment description, rubric and student submission:

{
  "criterion_scores": {},
  "criterion_status": {},
  "criterion_feedback": {},
  "overall_feedback": {}
}

The published tokenizer and configuration include a YaRN 8,192-token setting and stop on either <|endoftext|> (151643) or <|im_end|> (151645). This means inputs can be configured up to 8K tokens; it is not evidence of reliable long-context retrieval or assessment.

Results and limitations

On 50 reused development records, the selected checkpoint produced 82% strict JSON and 84% EOS completion. It also showed 100% severe repetition, only 2% exact criterion-ID sets, and a score MAE of 2.0 wherever numeric comparison was possible. It can therefore demonstrate a learned JSON-like feedback pattern, but is not a reliable autonomous assessor, educator replacement, or production Gemini replacement. The included training_run_report.json gives the exact aggregate run evidence; it contains no student submissions or raw predictions.

Load with Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "moinsaj/aaie-gft-full-fg-8k"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo, torch_dtype=torch.bfloat16, device_map="auto", attn_implementation="sdpa"
)

# `fg_prompt` must follow the AAIE FG prompt contract and ends with `Response:`.
messages = [{"role": "user", "content": fg_prompt}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
generated = model.generate(
    inputs, max_new_tokens=1600, do_sample=False,
    eos_token_id=[151643, 151645], pad_token_id=151643, use_cache=True,
)
answer = tokenizer.decode(generated[0, inputs.shape[1]:], skip_special_tokens=True)
print(answer)

Use deterministic decoding for reproducible diagnostic comparisons. Parse and validate the JSON before use; do not rely on the model to infer missing evidence or score an unverified student submission.

Training provenance

• Parent lineage: published AAIE GFT Llama-compatible checkpoint, revision 7dbc72d720237bd6abaaa3e7701ca69544a03e4a.
• Architecture: Llama causal LM, 20 layers, hidden size 512, 8 attention heads, 2 key/value heads, vocabulary 151,936.
• Fine-tuning: full weights, FP32 parameters with BF16 autocast, AdamW, learning rate 5e-6, 3% warm-up and cosine decay, batch 1 with accumulation 8, two epochs.
• Dataset: 1,621 fitting complete training records; 261 fitting validation records. evidence_map was retained in provenance only, not as generated content. No raw records, prompts, submissions, teacher feedback or predictions are published here.
• Checkpoint SHA-256: 95828700cfc649f8e2d0d75b400824e144855573f930871468cabd616b99689e.

Intended use

Research, reproducible testing and supervised demonstrations only. Human educator review is required for any feedback or assessment decision. Confirm the parent model's licence and your data-use obligations before reuse.